%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % % % % % A Bibliography on Automated Text Categorization % % % % updated and maintained by % % % % Evgeniy Gabrilovich % % Department of Computer Science % % Technion - Israel Institute of Technology % % Technion City, Haifa 32000, Israel % % Email: gabr AT cs . technion . ac . il % % WWW: http://www.cs.technion.ac.il/~gabr % % % % % % originally created by % % % % Fabrizio Sebastiani % % Dipartimento di Matematica Pura e Applicata % % Universita' di Padova % % Via Giovanni Battista Belzoni, 7 - 35131 Padova, Italy % % http://www.math.unipd.it/~fabseb60/ % % % % % % This is a bibliography, in BibTeX format, on automated text % % categorization (ATC), defined as the activity of automatically % % building, by means of machine learning techniques, automatic text % % classifiers, i.e. systems capable of assigning to a text % % document one or more thematic categories from a predefined set. % % % % This bibliography resides at % % http://www.cs.technion.ac.il/~gabr/resources/atc/ATCbibliography.bib % % A companion Web page is available at % % http://www.cs.technion.ac.il/~gabr/resources/atc/atcbib.html % % Everyone is welcome to download the bibliography as a whole and % % distribute it, provided that it is distributed untouched. % % % % Everyone is also welcome to let me know either additional % % references or corrections and additions (e.g. URLs, where % % they are not already present) to the existing ones. % % In general, only references specific to ATC are considered % % pertinent to this bibliography; in particular, references that % % *are* considered pertinent are: % % % % * publications that discuss novel ATC methods, novel % % experimentation of previously known methods, or resources for % % ATC experimentation; % % % % * publications that discuss applications of ATC (e.g. % % automated indexing for Boolean IR systems, filtering, etc.). % % % % References that are *not* considered pertinent are: % % % % * publications that discuss techniques in principle useful for % % ATC (e.g. machine learning techniques, information retrieval % % techniques) but do not explicitly discuss their application % % to ATC; % % % % * publications thet discuss related topics sometimes confused with % % ATC; these include, in particular, text clustering (i.e. text % % classification by unsupervised learning) and text indexing; % % % % * technical reports and workshop papers. Only papers that have % % been the object of formal publication (i.e. conferences and % % journals) are to be included in the bibliography, so as to avoid % % its explosion and the inclusion of material bound to obsolescence. % % % % Concerning URLs from which to download on-line copies of the % % papers, where possible I have included URLs with unrestricted % % access (e.g. home pages of authors). When such URLs were not % % available, sometimes a URL with restricted access (e.g. the % % ACM Digital Library or the IEEE Computing Society Digital % % Library, which are accessible to subscribers only) is indicated. % % When this is the case, if you know of a URL with unrestricted access % % from which the paper is also available, please let me know and I % % will substitute the link. % % % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% % % ENTRY TYPES % % Article % = an article from a journal or magazine % required: author, title, journal, year % optional: volume, number, pages, month, note % % Book % = a book with an explicit publisher % required: author or editor, title, publisher, year % optional: volume or number, series, address, edition, month, note % % Booklet % = a work that is printed and bound, but without a named publisher % or sponsoring institution % required: title % optional: author, howpublished, address, month, year, note % % Conference % = same as 'InProceedings', included for compatibility with older versions % % InBook % = a part of a book, usually untitled; it may be a chapter % (or other sectional unit) and/or range of pages % required: author or editor, title, chapter and/or pages, publisher, year % optional: volume or number, series, type, address, edition, month, note % % InCollection % = a part of a book with its own title % required: author, title, booktitle, publisher, year % optional: editor, volume or number, series, type, chapter, pages, % address, edition, month, note % % InProceedings % = an article in a conference proceedings % required: author, title, booktitle, year % optional: editor, volume or number, series, pages, address, month, % organization, publisher, note % % Manual % = technical documentation % required: title % optional: author, organization, address, edition, month, year, note % % MastersThesis % = a master's thesis % required: author, title, school, year % optional: type, address, month, note % % Misc % = use this type when nothing else fits % required: none % optional: author, title, howpublished, month, year, note % % PhDThesis % = a Ph.D. thesis % required: author, title, school, year % optional: type, address, month, note % % Proceedings % = the proceedings of a conference % required: title, year % optional: editor, volume or number, series, address, month, % organization, publisher, note % % TechReport % = a report published by a school or other institution, % usually numbered within a series % required: author, title, institution, year % optional: type, number, address, month, note % % Unpublished % = a document with an author and title, but not formally published % required: author, title, note % optional: month, year % %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% @inProceedings{Adam02, author = {Chai K. Adam and Hwee T. Ng and Hai L. Chieu}, title = {Bayesian Online Classifiers for Text Classification and Filtering}, booktitle = {Proceedings of SIGIR-02, 25th ACM International Conference on Research and Development in Information Retrieval}, editor = {Micheline Beaulieu and Ricardo Baeza-Yates and Sung Hyon Myaeng and Kalervo J{\"{a}}rvelin}, publisher = {ACM Press, New York, US}, address = {Tampere, FI}, year = {2002}, pages = {97--104}, url = {http://doi.acm.org/10.1145/564376.564395}, abstract = {This paper explores the use of Bayesian online classifiers to classify text documents. Empirical results indicate that these classifiers are comparable with the best text classification systems. Furthermore, the online approach offers the advantage of continuous learning in the batch-adaptive text filtering task.}, } @inProceedings{Adami03, author = {Giordano Adami and Paolo Avesani and Diego Sona}, title = {Bootstrapping for hierarchical document classification}, booktitle = {Proceedings of CIKM-03, 12th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {2003}, address = {New Orleans, US}, pages = {295--302}, url = {http://doi.acm.org/10.1145/956863.956920}, abstract = {Managing the hierarchical organization of data is starting to play a key role in the knowledge management community due to the great amount of human resources needed to create and maintain these organized repositories of information. Machine learning community has in part addressed this problem by developing hierarchical supervised classifiers that help maintainers to categorize new resources within given hierarchies. Although such learning models succeed in exploiting relational knowledge, they are highly demanding in terms of labeled examples, because the number of categories is related to the dimension of the corresponding hierarchy. Hence, the creation of new directories or the modification of existing ones require strong investments.This paper proposes a semi-automatic process (interleaved with human suggestions) whose aim is to minimize (simplify) the work required to the administrators when creating, modifying, and maintaining directories. Within this process, bootstrapping a taxonomy with examples represents a critical factor for the effective exploitation of any supervised learning model. For this reason we propose a method for the bootstrapping process that makes a first hypothesis of categorization for a set of unlabeled documents, with respect to a given empty hierarchy of concepts. Based on a revision of Self-Organizing Maps, namely TaxSOM, the proposed model performs an unsupervised classification, exploiting the a-priori knowledge encoded in a taxonomy structure both at the terminological and topological level. The ultimate goal of TaxSOM is to create the premise for successfully training a supervised classifier.}, } @inProceedings{Aggarwal99, author = {Charu C. Aggarwal and Stephen C. Gates and Philip S. Yu}, title = {On the merits of building categorization systems by supervised clustering}, booktitle = {Proceedings of EDBT-00, 7th International Conference on Extending Database Technology}, publisher = {ACM Press, New York, US}, year = {1999}, address = {Konstanz, DE}, pages = {352--356}, url = {http://doi.acm.org/10.1145/312129.312279}, abstract = {This paper investigates the use of supervised clustering in order to create sets of categories for classification of documents. We use information from a pre-existing taxonomy in order to supervise the creation of a set of related clusters, though with some freedom in defining and creating the classes. We show that the advantage of using supervised clustering is that it is possible to have some control over the range of subjects that one would like the categorization system to address, but with a precise mathematical definition of each category. We then categorize documents using this a priori knowledge of the definition of each category. We also discuss a new technique to help the classifier distinguish better among closely related clusters. Finally, we show empirically that this categorization system utilizing a machine-derived taxonomy performs as well as a manual categorization process, but at a far lower cost.}, } @inProceedings{Agrawal00, author = {Rakesh Agrawal and Roberto J. Bayardo and Ramakrishnan Srikant}, title = {{\sc Athena}: Mining-based Interactive Management of Text Databases}, booktitle = {Proceedings of EDBT-00, 7th International Conference on Extending Database Technulogy}, editor = {Carlo Zaniolo and Peter C. Lockemann and Marc H. Scholl and Torsten Grust}, year = {2000}, address = {Konstanz, DE}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1777}, pages = {365--379}, url = {http://www.almaden.ibm.com/cs/people/ragrawal/papers/athena.ps}, abstract = {We describe Athena: a system for creating, exploiting, and maintaining a hierarchical arrangement of textual documents through interactive mining-based operations. Requirements of any such system include speed and minimal end-user effort. Athena satisfies these requirements through linear-time classification and clustering engines which are applied interactively to speed the development of accurate models. Naive Bayes classifiers are recognized to be among the best for classifying text. We show that our specialization of the Naive Bayes classifier is considerably more accurate (7 to 29\% absolute increase in accuracy) than a standard implementation. Our enhancements include using Lidstone's law of succession instead of Laplace's law, under-weighting long documents, and over-weighting author and subject. We also present a new interactive clustering algorithm, C-Evolve, for topic discovery. C-Evolve first finds highly accurate cluster digests (partial clusters), gets user feedback to merge and correct these digests, and then uses the classification algorithm to complete the partitioning of the data. By allowing this interactivity in the clustering process, C-Evolve achieves considerably higher clustering accuracy (10 to 20\% absolute increase in our experiments) than the popular K-Means and agglomerative clustering methods.}, } @inProceedings{Agrawal01, author = {Rakesh Agrawal and Ramakrishnan Srikant}, title = {On integrating catalogs}, booktitle = {Proceedings of WWW-01, 10th International Conference on the World Wide Web}, publisher = {ACM Press, New York, US}, editor = {}, year = {2001}, address = {Hong Kong, CN}, pages = {603--612}, url = {http://doi.acm.org/10.1145/371920.372163}, abstract = {We address the problem of integrating documents from different sources into a master catalog. This problem is pervasive in web marketplaces and portals. Current technology for automating this process consists of building a classifier that uses the categorization of documents in the master catalog to construct a model for predicting the category of unknown documents. Our key insight is that many of the data sources have their own categorization, and classification accuracy can be improved by factoring in the implicit information in these source categorizations. We show how a Naive Bayes classification can be enhanced to incorporate the similarity information present in source catalogs. Our analysis and empirical evaluation show substantial improvement in the accuracy of catalog integration.}, } @inProceedings{Aizawa00, author = {Akiko Aizawa}, title = {The feature quantity: an information-theoretic perspective of tfidf-like measures}, booktitle = {Proceedings of SIGIR-00, 23rd ACM International Conference on Research and Development in Information Retrieval}, editor = {Nicholas J. Belkin and Peter Ingwersen and Mun-Kew Leong}, publisher = {ACM Press, New York, US}, address = {Athens, GR}, year = {2000}, pages = {104--111}, url = {http://doi.acm.org/10.1145/345508.345556}, abstract = {The feature quantity, a quantitative representation of specificity introduced in this paper, is based on an information theoretic perspective of co-occurrence events between terms and documents. Mathematically, the feature quantity is defined as a product of probabillty and information, and maintains a good correspondence with the tfidf-like measures popularly used in today's IR systems. In this paper, we present a formal description of the feature quantity, as well as some illustrative examples of applying such a quantity to different types of information retrieval tasks: representative term selection and text categorization.}, } @inProceedings{Aizawa01, author = {Akiko Aizawa}, title = {Linguistic Techniques to Improve the Performance of Automatic Text Categorization}, booktitle = {Proceedings of NLPRS-01, 6th Natural Language Processing Pacific Rim Symposium}, editor = {}, publisher = {}, address = {Tokyo, JP}, year = {2001}, pages = {307--314}, url = {http://www.afnlp.org/nlprs2001/pdf/0079-01.pdf}, abstract = {This paper presents a method for incorporating natural language processing into existing text categorization procedures. Three aspects are considered in the investigation: (i) a method for weighting terms based on the concept of a probability weighted amount of information, (ii) estimation of term occurrence probabilities using a probabilistic language model, and (iii) automatic extraction of terms based on POS tags automatically generated by a morphological analyzer. The effects of these considerations are examined in the experiments using Reuters-21578 and NTCIR-J1 standard test collections.}, } @inProceedings{Alias02, author = {Francesc Al{\'i}as and Ignasi Iriondo and Pere Barnola}, title = {Multi-domain text classification for unit selection text-to-speech synthesis}, booktitle = {Proceedings of ICPhS-03, 15th International Congress on Phonetic Sciences}, address = {Barcelona, ES}, editor = {}, publisher = {}, year = {2003}, pages = {}, url = {}, abstract = {}, } @inProceedings{AlKofahi01, author = {Khalid Al-Kofahi and Alex Tyrrell and Arun Vachher and Tim Travers and Peter Jackson}, title = {Combining Multiple Classifiers for Text Categorization}, booktitle = {Proceedings of CIKM-01, 10th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {Henrique Paques and Ling Liu and David Grossman}, year = {2001}, address = {Atlanta, US}, pages = {97--104}, url = {http://doi.acm.org/10.1145/502585.502603}, abstract = {A major problem facing online information services is how to index and supplement large document collections with respect to a rich set of categories. We focus upon the routing of case law summaries to various secondary law volumes in which they should be cited. Given the large number (> 13,000) of closely related categories, this is a challenging task that is unlikely to succumb to a single algorithmic solution. Our fully implemented and recently deployed system shows that a superior classification engine for this task can be constructed from a combination of classifiers. The multi-classifier approach helps us leverage all the relevant textual features and meta data, and appears to generalize to related classification tasks.}, } @inProceedings{Amati96, author = {Gianni Amati and Daniela D'Aloisi and Vittorio Giannini and Flavio Ubaldini}, title = {An Integrated System for Filtering News and Managing Distributed Data}, booktitle = {Proceedings of PAKM-96, 1st International Conference on Practical Aspects of Knowledge Management}, editor = {}, publisher = {}, year = {1996}, pages = {}, note = {An extended version appears as~\cite{Amati97b}}, address = {Basel, CH}, url = {http://airone.fub.it:8080/projects/pakm96.ps}, abstract = {With the development and diffusion of the Internet worldwide connection, a large amount of information can be delivered to the users. To avoid their being overflowed by the incoming data, methods of information filtering are required. Thus, there is the problem of determining what information is relevant to the user and how this decision can be taken by a supporting system. Parametric and qualitative descriptors of user's interest must be generated. This paper presents two approaches. The first concerns an information filtering system based on an adaptation of the generalized probabilistic model of information retrieval. The user profile is a vector of weighted terms which are learned from the relevance assessment values given by the user on the training set. Positive terms are considered relevant to the informative need of the user, negative ones irrelevant. The relevance values are interpreted as subjective probabilities and hence are mapped into the real interval [0; 1]. ProFile is a filtering system for the netnews which uses this model with a scale of 11 predefined values of relevance. ProFile allows the user to update on-line his profile and to check the discrepancy between his assessment and the prediction of relevance of the system. The second concerns the InfoAgent, a system for supporting users in retrieving data in distributed and heterogeneous archives and repositories. The architecture is based on the metaphor of the software agents and incorporates innovative hints from other fields: distributed architectures, relevance feedback and active interfaces. The system has a cooperative and supportive role: it understands the user's needs and learns from his behavior. Its aim is to disengage the user from learning complex tools and from performing tedious and repetitive actions.}, } @inProceedings{Amati97, author = {Gianni Amati and Fabio Crestani and Flavio Ubaldini}, title = {A learning system for selective dissemination of information}, booktitle = {Proceedings of IJCAI-97, 15th International Joint Conference on Artificial Intelligence}, editor = {Martha E. Pollack}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, year = {1997}, pages = {764--769}, address = {Nagoya, JP}, url = {http://www.cs.strath.ac.uk/~fabioc/papers/97-ijcai.pdf}, abstract = {New methods and new systems are needed to filter or to selectively distribute the increasing volume of electronic information being produced nowadays. An effective information filtering system is one that provides the exact information that fulfills a user's interest with the minimum effort by the user to describe it. Such a system will have to be adaptive to the user changing interest. In this paper we present a learning system for information filtering and selective information dissemination. The learning algorithm is described and the effectiveness of the system is evaluated in a true information filtering style.}, } @inProceedings{Amati97a, author = {Gianni Amati and Fabio Crestani and Flavio Ubaldini and Stefano De Nardis}, title = {Probabilistic Learning for Information Filtering}, booktitle = {Proceedings of RIAO-97, 1st International Conference ``Recherche d'Information Assistee par Ordinateur''}, editor = {Luc Devroye and Claude Chrisment}, address = {Montreal, CA}, year = {1997}, pages = {513--530}, note = {An extended version appears as~\cite{Amati99}}, url = {http://www.cs.strath.ac.uk/~fabioc/papers/97-riao.pdf}, abstract = {In this paper we describe and evaluate a learning model for information filtering which is an adaptation of the generalised probabilistic model of Information Retrieval. The model is based on the concept of ``uncertainty sampling'', a technique that allows for relevance feedback both on relevant and non relevant documents. The proposed learning model is the core of a prototype information filtering system called ProFile.}, } @article{Amati97b, author = {Gianni Amati and Daniela D'Aloisi and Vittorio Giannini and Flavio Ubaldini}, title = {A Framework for Filtering News and Managing Distributed Data}, journal = {Journal of Universal Computer Science}, year = {1997}, number = {8}, volume = {3}, pages = {1007--1021}, url = {http://www.jucs.org/jucs_3_8/a_framework_for_filtering}, abstract = {With the development and diffusion of the Internet worldwide connection, a large amount of information is available to the users. Methods of information filtering and fetching are then required. This paper presents two approaches. The first concerns the information filtering system ProFile based on an adaptation of the generalized probabilistic model of information retrieval. ProFile filters the netnews and uses a scale of 11 predefined values of relevance. ProFile allows the user to update on-line the profile and to check the discrepancy between the assessment and the prediction of relevance of the system. The second concerns ABIS, an intelligent agent for supporting users in filtering data from distributed and heterogeneous archives and repositories. ABIS minimizes user's effort in selecting the huge amount of available documents. The filtering engine memorizes both user preferences and past situations. ABIS compares documents with the past situations and finds the similarity scores on the basis of a memory-based reasoning approach.}, } @article{Amati99, author = {Gianni Amati and Fabio Crestani}, title = {Probabilistic learning for selective dissemination of information}, journal = {Information Processing and Management}, pages = {633--654}, year = {1999}, number = {5}, volume = {35}, url = {http://www.cs.strath.ac.uk/~fabioc/papers/99-ipem.pdf}, abstract = {New methods and new systems are needed to filter or to selectively distribute the increasing volume of electronic information being produced nowadays. An effective information filtering system is one that provides the exact information that fulfills user's interests with the minimum effort by the user to describe it. Such a system will have to be adaptive to the user changing interest. In this paper we describe and evaluate a learning model for information filtering which is an adaptation of the generalized probabilistic model of Information Retrieval. The model is based on the concept of `uncertainty sampling', a technique that allows for relevance feedback both on relevant and nonrelevant documents. The proposed learning model is the core of a prototype information filtering system called ProFile.}, } @inProceedings{Androutsopoulos00, author = {Ion Androutsopoulos and John Koutsias and Konstandinos V. Chandrinos and Constantine D. Spyropoulos}, title = {An experimental comparison of naive Bayesian and keyword-based anti-spam filtering with personal e-mail messages}, booktitle = {Proceedings of SIGIR-00, 23rd ACM International Conference on Research and Development in Information Retrieval}, editor = {Nicholas J. Belkin and Peter Ingwersen and Mun-Kew Leong}, publisher = {ACM Press, New York, US}, address = {Athens, GR}, year = {2000}, pages = {160--167}, url = {http://doi.acm.org/10.1145/345508.345569}, abstract = {The growing problem of unsolicited bulk e-mail, also known as ``spam'', has generated a need for reliable anti-spam e-mail filters. Filters of this type have so far been based mostly on manually constructed keyword patterns. An alternative approach has recently been proposed, whereby a Naive Bayesian classifier is trained automatically to detect spam messages. We test this approach on a large collection of personal e-mail messages, which we make publicly available in "encrypted" form contributing towards standard benchmarks. We introduce appropriate cost-sensitive measures, investigating at the same time the effect of attribute-set size, training-corpus size, lemmatization, and stop lists, issues that have not been explored in previous experiments. Finally, the Naive Bayesian filter is compared, in terms of performance, to a filter that uses keyword patterns, and which is part of a widely used e-mail reader.}, } @article{Appiani01, author = {Enrico Appiani and Francesca Cesarini and Annamaria Colla and Massimiliano Diligenti and Marco Gori and Simone Marinai and Giovanni Soda}, title = {Automatic document classification and indexing in high-volume applications}, journal = {International Journal on Document Analysis and Recognition}, year = {2001}, number = {2}, volume = {4}, pages = {69--83}, url = {http://link.springer-ny.com/link/service/journals/10032/papers/1004002/10040069.pdf}, abstract = {In this paper a system for analysis and automatic indexing of imaged documents for high-volume applications is described. This system, named STRETCH (STorage and RETrieval by Content of imaged documents), is based on an Archiving and Retrieval Engine, which overcomes the bottleneck of document profiling bypassing some limitations of existing pre-defined indexing schemes. The engine exploits a structured document representation and can activate appropriate methods to characterise and automatically index heterogeneous documents with variable layout. The originality of STRETCH lies principally in the possibility for unskilled users to define the indexes relevant to the document domains of their interest by simply presenting visual examples and applying reliable automatic information extraction methods (document classification, flexible reading strategies) to index the documents automatically, thus creating archives as desired. STRETCH offers ease of use and application programming and the ability to dynamically adapt to new types of documents. The system has been tested in two applications in particular, one concerning passive invoices and the other bank documents. In these applications, several classes of documents are involved. The indexing strategy first automatically classifies the document, thus avoiding pre-sorting, then locates and reads the information pertaining to the specific document class. Experimental results are encouraging overall; in particular, document classification results fulfill the requirements of high-volume application. Integration into production lines is under execution.}, } @article{Apte94, author = {Apt\'{e}, Chidanand and Damerau, Fred J. and Weiss, Sholom M.}, title = {Automated learning of decision rules for text categorization}, journal = {ACM Transactions on Information Systems}, year = {1994}, number = {3}, volume = {12}, pages = {233--251}, url = {http://www.acm.org/pubs/articles/journals/tois/1994-12-3/p233-apte/p233-apte.pdf}, abstract = {We describe the results of extensive experiments using optimized rule-based induction methods on large document collections. The goal of these methods is to discover automatically classification patterns that can be used for general document categorization or personalized filtering of free text. Previous reports indicate that human-engineered rule-based systems, requiring many man-years of developmental efforts, have been successfully built to ``read'' documents and assign topics to them. We show that machine-generated decision rules appear comparable to human performance, while using the identical rule-based representation. In comparison with other machine-learning techniques, results on a key benchmark from the Reuters collection show a large gain in performance, from a previously reported 67\% recall/precision breakeven point to 80.5\%. In the context of a very high-dimensional feature space, several methodological alternatives are examined, including universal versus local dictionaries, and binary versus frequency related features.}, } @inProceedings{Apte94a, author = {Apt\'{e}, Chidanand and Damerau, Fred J. and Weiss, Sholom M.}, title = {Towards Language-Independent Automated Learning of Text Categorization Models}, booktitle = {Proceedings of SIGIR-94, 17th ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and Van Rijsbergen, Cornelis J.}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Dublin, IE}, pages = {23--30}, year = {1994}, note = {An extended version appears as~\cite{Apte94}}, url = {http://www.acm.org/pubs/articles/proceedings/ir/188490/p23-apte/p23-apte.pdf}, abstract = {We describe the results of extensive machine learning experiments on large collections of Reuters' English and German newswires. The goal of these experiments was to automatically discover classification patterns that can be used for assignment of topics to the individual newswires. Our results with the English newswire collection show a very large gain in performance as compared to published benchmarks, while our initial results with the German newswires appear very promising. We present our methodology, which seems to be insensitive to the language of the document collections, and discuss issues related to the differences in results that we have obtained for the two collections.}, } @article{Attardi98, author = {Attardi, Giuseppe and Di Marco, Sergio and Salvi, Davide}, title = {Categorization by context}, journal = {Journal of Universal Computer Science}, year = {1998}, number = {9}, volume = {4}, pages = {719--736}, url = {http://www.jucs.org/jucs_4_9/categorisation_by_context}, abstract = {Assistance in retrieving of documents on the World Wide Web is provided either by search engines, through keyword based queries, or by catalogues, which organise documents into hierarchical collections. Maintaining catalogues manually is becoming increasingly difficult due to the sheer amount of material on the Web, and therefore it will be soon necessary to resort to techniques for automatic classification of documents. Classification is traditionally performed by extracting information for indexing a document from the document itself. The paper describes the technique of categorisation by context, which exploits the context perceivable from the structure of HTML documents to extract useful information for classifying the documents they refer to. We present the results of experiments with a preliminary implementation of the technique.}, } @inProceedings{Attardi99, author = {Giuseppe Attardi and Antonio Gull{\'{\i}} and Fabrizio Sebastiani}, title = {Automatic Web Page Categorization by Link and Context Analysis}, booktitle = {Proceedings of THAI-99, 1st European Symposium on Telematics, Hypermedia and Artificial Intelligence}, editor = {Chris Hutchison and Gaetano Lanzarone}, year = {1999}, address = {Varese, IT}, pages = {105--119}, url = {http://www.math.unipd.it/~fabseb60/Publications/THAI99.pdf}, abstract = {Assistance in retrieving documents on the World Wide Web is provided either by search engines, through keyword-based queries, or by catalogues, which organize documents into hierarchical collections. Maintaining catalogues manually is becoming increasingly difficult, due to the sheer amount of material on the Web; it is thus becoming necessary to resort to techniques for the automatic classification of documents. Automatic classification is traditionally performed by extracting the information for representing a document (``indexing'') from the document itself. The paper describes the novel technique of categorization by context, which instead extracts useful information for classifying a document from the context where a URL referring to it appears. We present the results of experimenting with Theseus, a classifier that exploits this technique.}, } @inProceedings{Avancini03, author = {Henri Avancini and Alberto Lavelli and Bernardo Magnini and Fabrizio Sebastiani and Roberto Zanoli}, title = {Expanding Domain-Specific Lexicons by Term Categorization}, year = {2003}, booktitle = {Proceedings of SAC-03, 18th ACM Symposium on Applied Computing}, address = {Melbourne, US}, publisher = {ACM Press, New York, US}, pages = {793--797}, url = {http://www.math.unipd.it/~fabseb60/Publications/SAC03c.pdf}, abstract = {We discuss an approach to the automatic expansion of domain-specific lexicons by means of \emph{term categorization}, a novel task employing techniques from information retrieval (IR) and machine learning (ML). Specifically, we view the expansion of such lexicons as a process of learning previously unknown associations between terms and \emph{domains}. The process generates, for each $c_{i}$ in a set $C=\{c_{1},\ldots,c_{m}\}$ of domains, a lexicon $L^{i}_{1}$, bootstrapping from an initial lexicon $L^{i}_{0}$ and a set of documents $\theta$ given as input. The method is inspired by \emph{text categorization} (TC), the discipline concerned with labelling natural language texts with labels from a predefined set of domains, or categories. However, while TC deals with documents represented as vectors in a space of terms, we formulate the task of term categorization as one in which terms are (dually) represented as vectors in a space of documents, and in which terms (instead of documents) are labelled with domains.}, } @inProceedings{Baker98, author = {Douglas Baker and Andrew K. McCallum}, title = {Distributional clustering of words for text classification}, booktitle = {Proceedings of the 21st ACM International Conference on Research and Development in Information Retrieval}, editor = {Bruce Croft and Alistair Moffat and Van Rijsbergen, Cornelis J. and Ross Wilkinson and Justin Zobel}, publisher = {ACM Press, New York, US}, year = {1998}, address = {Melbourne, AU}, pages = {96--103}, url = {http://www.cs.cmu.edu/~mccallum/papers/clustering-sigir98.ps.gz}, abstract = {We describe the application of distributional clustering to document classification. This approach clusters words into groups based on the distribution of class labels associated with each word. Thus, unlike some other unsupervised dimensionality-reduction techniques, such as latent semantic indexing, we are able to compress the feature space much more aggressively, while still maintaining high document classification accuracy. Experimental results obtained on three real-world data sets show that we can reduce the feature dimensionality by three orders of magnitude and lose only 2\% accuracy, significantly better than latent semantic indexing, class-based clustering, feature selection by mutual information, or Markov-blanket-based feature selection. We also show that less aggressive clustering sometimes results in improved classification accuracy over classification without clustering.}, } @inProceedings{Bao01, author = {Yongguang Bao and Satoshi Aoyama and Xiaoyong Du and Kazutaka Yamada and Naohiro Ishii}, title = {A Rough Set-Based Hybrid Method to Text Categorization}, booktitle = {Proceedings of WISE-01, 2nd International Conference on Web Information Systems Engineering}, editor = {M. Tamer {\"O}zsu and Hans-J{\"{o}}rg Schek and Katsumi Tanaka and Yanchun Zhang and Yahiko Kambayashi}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, year = {2001}, address = {Kyoto, JP}, pages = {254--261}, url = {}, abstract = {In this paper we present a hybrid text categorization method based on Rough Sets theory. A central problem in good text Classification for information filtering and retrieval (IF/IR) is the high dimensionality of the data. It may contain many unnecessary and irrelevant features. To cope with this problem, we propose a hybrid technique using Latent Semantic Indexing (LSI) and Rough Sets theory (RS) to alleviate this situation. Given corpora of documents and a training set of examples of classified documents, the technique locates a minimal set of co-ordinate keywords to distinguish between classes of documents, reducing the dimensionality of the keyword vectors. This simplifies the creation of knowledge-based IF/IR systems, speeds up their operation, and allows easy editing of the rule bases employed. Besides, we generate several knowledge base instead of one knowledge base for the classification of new object, hoping that the combination of answers of the multiple knowledge bases result in better performance. Multiple knowledge bases can be formulated precisely and in a unified way within the framework of RS. This paper describes the proposed technique, discusses the integration of a keyword acquisition algorithm, Latent Semantic Indexing (LSI) with Rough Set-based rule generate algorithm, and provides experimental results. The test results show the hybrid method is better than the previous rough set-based approach.}, } @inProceedings{Basili00, author = {Roberto Basili and Alessandro Moschitti and Maria T. Pazienza}, title = {Language-Sensitive Text Classification}, booktitle = {Proceedings of RIAO-00, 6th International Conference ``Recherche d'Information Assistee par Ordinateur''}, editor = {}, address = {Paris, France}, year = {2000}, pages = {331--343}, url = {}, abstract = {It is a traditional belief that in order to scale-up to more effective retrieval and access methods modern Information Retrieval has to consider more the text content. The modalities and techniques to fit this objectives are still under discussion. More empirical evidence is required to determine the suitable linguistic levels for modeling each IR subtask (e.g. information zoning, parsing, feature selection for indexing,...) and the corresponding use of this information. In this paper an original classification model sensitive to document syntactic information and characterized by a novel inference method is described. Extensive experimental evidence has been derived on real test data and also from well-established academic test sets. The results show that a significant improvement can be derived using the proposed inference model. Also the role of linguistic preprocessing seems to provide positive effects on the performance. POS tagging and recognition of Proper Nouns received a specific experimental attention and provided significant effects on measured accuracy.}, } @inProceedings{Basili01, author = {Roberto Basili and Alessandro Moschitti and Maria T. Pazienza}, title = {NLP-driven IR: Evaluating Performances over a Text Classification task}, booktitle = {Proceeding of IJCAI-01, 17th International Joint Conference on Artificial Intelligence}, editor = {Bernhard Nebel}, address = {Seattle, US}, year = {2001}, pages = {1286--1291}, url = {}, abstract = {Although several attempts have been made to introduce Natural Language Processing (NLP) techniques in Information Retrieval, most ones failed to prove their effectiveness in increasing performances. In this paper Text Classification (TC) has been taken as the IR task and the effect of linguistic capabilities of the underlying system have been studied. A novel model for TC, extending a well know statistical model (i.e. Rocchio's formula [Ittner et al., 1995]) and applied to linguistic features has been defined and experimented. The proposed model represents an effective feature selection methodology. All the experiments result in a significant improvement with respect to other purely statistical methods (e.g. [Yang, 1999]), thus stressing the relevance of the available linguistic information. Moreover, the derived classifier reachs the performance (about 85\%) of the best known models (i.e. Support Vector Machines (SVM) and k-Nearest Neighbour (KNN)) characterized by an higher computational complexity for training and processing.}, } @inProceedings{Basili01a, author = {Roberto Basili and Alessandro Moschitti and Maria T. Pazienza}, title = {An hybrid approach to optimize feature selection process in text classification}, booktitle = {Proceedings of AI*IA-01, 7th Congress of the Italian Association for Artificial Intelligence}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2175}, editor = {Floriana Esposito}, year = {2001}, pages = {320--325}, address = {Bari, IT}, url = {http://link.springer.de/link/service/series/0558/papers/2175/21750320.pdf}, abstract = {Feature selection and weighting are the primary activity of every learning algorithm for text classification. Traditionally these tasks are carried out individually in two distinct phases: the first is the global feature selection during a corpus pre-processing and the second is the application of the feature weighting model. This means that two (or several) different techniques are used to optimize the performances even if a single algorithm may have more chances to operate the right choices. When the complete feature set is available, the classifier learning algorithm can better relate to the suitable representation level the different complex features like linguistic ones (e.g. syntactic categories associated to words in the training material or terminological expressions). In [3] it has been suggested that classifiers based on generalized Rocchio formula can be used to weight features in category profiles in order to exploit the selectivity of linguistic information techniques in text classification. In this paper, a systematic study aimed to understand the role of Rocchio formula in selection and weighting of linguistic features will be described.}, } @inProceedings{Basili01b, author = {Roberto Basili and Alessandro Moschitti}, title = {A robust model for intelligent text classification}, booktitle = {Proceedings of ICTAI-01, 13th IEEE International Conference on Tools with Artificial Intelligence}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, editor = {}, year = {2001}, pages = {265--272}, address = {Dallas, US}, url = {}, abstract = {Methods for taking into account linguistic content into text retrieval are receiving a growing attention [16],[14]. Text categorization is an interesting area for evaluating and quantifying the impact of linguistic information. Works in text retrieval through Internet suggest that embedding linguistic information at a suitable level within traditional quantitative approaches (e.g. sense distinctions for query expansion as in [14]) is the crucial issue able to bring the experimental stage to operational results. This kind of representational problem is also studied in this paper where traditional methods for statistical text categorization are augmented via a systematic use of linguistic information. Again, as in [14], the addition of NLP capabilities also suggested a different application of existing methods in revised forms. This paper presents an extension of the Rocchio formula [11] as a feature weighting and selection model used as a basis for multilingual Information Extraction. It allows an effective exploitation of the available linguistic information that better emphasizes this latter with significant both data compression and accuracy. The results is an original statistical classifier fed with linguistic (i.e. more complex) features and characterized by the novel feature selection and weighting model. It outperforms existing systems by keeping most of their interesting properties (i.e. easy implementation, low complexity and high scalability). Extensive tests of the model suggest its application as a viable and robust tool for large scale text classification and filtering, as well as a basic module for more complex scenarios.}, } @article{Bayer98, author = {Thomas Bayer and Ulrich Kressel and Heike Mogg-Schneider and Ingrid Renz}, title = {Categorizing paper documents. A generic system for domain and language independent text categorization}, journal = {Computer Vision and Image Understanding}, year = {1998}, number = {3}, volume = {70}, pages = {299--306}, url = {http://www.idealibrary.com/links/doi/10.1006/cviu.1998.0687/pdf}, abstract = {Text categorization assigns predefined categories to either electronically available texts or those resulting from document image analysis. A generic system for text categorization is presented which is based on statistical analysis of representative text corpora. Significant features are automatically derived from training texts by selecting substrings from actual word forms and applying statistical information and general linguistic knowledge. The dimension of the feature vectors is then reduced by linear transformation, keeping the essential information. The classification is a minimum least-squares approach based on polynomials. The described system can be efficiently adapted to new domains or different languages. In application, the adapted text categorizers are reliable, fast, and completely automatic. Two example categorization tasks achieve recognition scores of approximately 80\% and are very robust against recognition or typing errors.}, } @inProceedings{Bekkerman01, author = {Ron Bekkerman and Ran El-Yaniv and Naftali Tishby and Yoad Winter}, title = {On Feature Distributional Clustering for Text Categorization}, booktitle = {Proceedings of SIGIR-01, 24th ACM International Conference on Research and Development in Information Retrieval}, editor = {Croft, W. Bruce and Harper, David J. and Kraft, Donald H. and Zobel, Justin}, publisher = {ACM Press, New York, US}, address = {New Orleans, US}, year = {2001}, pages = {146--153}, url = {http://www.cs.huji.ac.il/labs/learning/Papers/sigir.ps.gz}, abstract = {We describe a text categorization approach that is based on a combination of feature distributional clusters with a support vector machine (SVM) classifier. Our feature selection approach employs distributional clustering of words via the recently introduced information bottleneck method, which generates a more efficient word-cluster representation of documents. Combined with the classification power of an SVM, this method yields high performance text categorization that can outperform other recent methods in terms of categorization accuracy and representation efficiency. Comparing the accuracy of our method with other techniques, we observe significant dependency of the results on the data set. We discuss the potential reasons for this dependency.}, } @article{Bekkerman03, author = {Ron Bekkerman and Ran El-Yaniv and Naftali Tishby and Yoad Winter}, title = {Distributional word clusters vs.\ words for text categorization}, journal = {Journal of Machine Learning Research}, volume = {3}, pages = {1183--1208}, year = {2003}, url = {http://www.jmlr.org/papers/volume3/bekkerman03a/bekkerman03a.pdf}, abstract = {We study an approach to text categorization that combines distributional clustering of words and a Support Vector Machine (SVM) classifier. This word-cluster representation is computed using the recently introduced Information Bottleneck method, which generates a compact and efficient representation of documents. When combined with the classification power of the SVM, this method yields high performance in text categorization. This novel combination of SVM with word-cluster representation is compared with SVM-based categorization using the simpler bag-of-words (BOW) representation. The comparison is performed over three known datasets. On one of these datasets (the 20 Newsgroups) the method based on word clusters significantly outperforms the word-based representation in terms of categorization accuracy or representation efficiency. On the two other sets (Reuters-21578 and WebKB) the word-based representation slightly outperforms the word-cluster representation. We investigate the potential reasons for this behavior and relate it to structural differences between the datasets.}, } @inProceedings{Bel03, author = {Nuria Bel and Cornelis H. Koster and Marta Villegas}, title = {Cross-lingual text categorization}, booktitle = {Proceedings of ECDL-03, 7th European Conference on Research and Advanced Technology for Digital Libraries}, editor = {Traugott Koch and Torvik S{\o}lvberg, Ingeborg}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2769}, year = {2003}, address = {Trondheim, NO}, pages = {126--139}, url = {}, abstract = {}, } @article{Benkhalifa01, author = {Mohammed Benkhalifa and Abdelhak Mouradi and Houssaine Bouyakhf}, title = {Integrating External Knowledge to Supplement Training Data in Semi-Supervised Learning for Text Categorization}, journal = {Information Retrieval}, number = {2}, volume = {4}, pages = {91--113}, year = {2001}, url = {http://www.wkap.nl/article.pdf?351286}, abstract = {Text Categorization (TC) is the automated assignment of text documents to predefined categories based on document contents. TC has been an application for many learning approaches, which prove effective. Nevertheless, TC provides many challenges to machine learning. In this paper, we suggest, for text categorization, the integration of external WordNet lexical information to supplement training data for a semi-supervised clustering algorithm which can learn from both training and test documents to classify new unseen documents. This algorithm is the ``Semi-Supervised Fuzzy c-Means'' (ssFCM). Our experiments use Reuters 21578 database and consist of binary classifications for categories selected from the 115 TOPICS classes of the Reuters collection. Using the Vector Space Model, each document is represented by its original feature vector augmented with external feature vector generated using WordNet. We verify experimentally that the integration of WordNet helps ssFCM improve its performance, effectively addresses the classification of documents into categories with few training documents and does not interfere with the use of training data.}, } @article{Benkhalifa01a, author = {Mohammed Benkhalifa and Abdelhak Mouradi and Houssaine Bouyakhf}, title = {Integrating {WordNet} knowledge to supplement training data in semi-supervised agglomerative hierarchical clustering for text categorization}, journal = {International Journal of Intelligent Systems}, pages = {929--947}, year = {2001}, volume = {16}, number = {8}, url = {http://www3.interscience.wiley.com/cgi-bin/fulltext?ID=84503376&PLACEBO=IE.pdf}, abstract = {The text categorization (TC) is the automated assignment of text documents to predefined categories based on document contents. TC has been an application for many learning approaches. which proved effective. Nevertheless, TC provides many challenges to machine learning. In this paper. we suggest, for text categorization, the integration of external WordNet lexical information to supplement training data for a semi-supervised clustering algorithm which (i) uses a finite design set of labeled data to (ii) help agglomerative hierarchical clustering algorithms (AHC) partition a finite set of unlabeled data and then (iii) terminates without the capacity to classify other objects. This algorithm is the "semi-supervised agglomerative hierarchical clustering algorithm" (ssAHC). Our experiments use Reuters 21578 database and consist of binary classifications for categories selected from the 89 TOPICS classes of the Reuters collection. Using the vector space model (VSM), each document is represented by its original feature vector augmented with external feature vector generated using WordNet. We verify experimentally that the integration of WordNet helps ssAHC improve its performance, effectively addresses the classification of documents into categories with few training documents. and does not interfere with the use of training data.}, } @inProceedings{Benkhalifa99, author = {Benkhalifa, Mohamed and Bensaid, Amine and Mouradi, Abdelhak}, title = {Text categorization using the semi-supervised fuzzy c-means algorithm}, booktitle = {Proceedings of NAFIPS-99, 18th International Conference of the North American Fuzzy Information Processing Society}, address = {New York, US}, pages = {561--565}, year = {1999}, url = {}, abstract = {Text categorization (TC) is the automated assignment of text documents to predefined categories based on document contents. TC has become very important in the information retrieval area, where information needs have tremendously increased with the rapid growth of textual information sources such as the Internet. We compare, for text categorization, two partially supervised (or semi-supervised) clustering algorithms: the Semi-Supervised Agglomerative Hierarchical Clustering (ssAHC) algorithm (A. Amar et al., 1997) and the Semi-Supervised Fuzzy-c-Means (ssFCM) algorithm (M. Amine et al., 1996). This (semi-supervised) learning paradigm falls somewhere between the fully supervised and the fully unsupervised learning schemes, in the sense that it exploits both class information contained in labeled data (training documents) and structure information possessed by unlabeled data (test documents) in order to produce better partitions for test documents. Our experiments, make use of the Reuters 21578 database of documents and consist of a binary classification for each of the ten most populous categories of the Reuters database. To convert the documents into vector form, we experiment with different numbers of features, which we select, based on an information gain criterion. We verify experimentally that ssFCM both outperforms and takes less time than the Fuzzy-c-Means (FCM) algorithm. With a smaller number of features, ssFCM's performance is also superior to that of ssAHC's. Finally ssFCM results in improved performance and faster execution time as more weight is given to training documents.}, } @inProceedings{Bennett02, author = {Paul N. Bennett and Susan T. Dumais and Eric Horvitz}, title = {Probabilistic combination of text classifiers using reliability indicators: models and results}, booktitle = {Proceedings of SIGIR-02, 25th ACM International Conference on Research and Development in Information Retrieval}, editor = {Micheline Beaulieu and Ricardo Baeza-Yates and Sung Hyon Myaeng and Kalervo J{\"{a}}rvelin}, publisher = {ACM Press, New York, US}, address = {Tampere, FI}, year = {2002}, pages = {207--214}, url = {http://doi.acm.org/10.1145/564376.564413}, abstract = {The intuition that different text classifiers behave in qualitatively different ways has long motivated attempts to build a better metaclassifier via some combination of classifiers. We introduce a probabilistic method for combining classifiers that considers the context-sensitive reliabilities of contributing classifiers. The method harnesses reliability indicators---variables that provide a valuable signal about the performance of classifiers in different situations. We provide background, present procedures for building metaclassifiers that take into consideration both reliability indicators and classifier outputs, and review a set of comparative studies undertaken to evaluate the methodology.}, } @inProceedings{Bennett03, author = {Paul N. Bennett}, title = {Using asymmetric distributions to improve text classifier probability estimates}, booktitle = {Proceedings of SIGIR-03, 26th ACM International Conference on Research and Development in Information Retrieval}, editor = {Jamie Callan and Gordon Cormack and Charles Clarke and David Hawking and Alan Smeaton}, publisher = {ACM Press, New York, US}, address = {Toronto, CA}, year = {2003}, pages = {111--118}, url = {http://doi.acm.org/10.1145/860435.860457}, abstract = {Text classifiers that give probability estimates are more readily applicable in a variety of scenarios. For example, rather than choosing one set decision threshold, they can be used in a Bayesian risk model to issue a run-time decision which minimizes a user-specified cost function dynamically chosen at prediction time. However, the quality of the probability estimates is crucial. We review a variety of standard approaches to converting scores (and poor probability estimates) from text classifiers to high quality estimates and introduce new models motivated by the intuition that the empirical score distribution for the "extremely irrelevant", "hard to discriminate", and "obviously relevant" items are often significantly different. Finally, we analyze the experimental performance of these models over the outputs of two text classifiers. The analysis demonstrates that one of these models is theoretically attractive (introducing few new parameters while increasing flexibility), computationally efficient, and empirically preferable.}, } @article{Bennett05, author = {Paul N. Bennett and Susan T. Dumais and Eric Horvitz}, title = {The Combination of Text Classifiers Using Reliability Indicators}, journal = {Information Retrieval}, number = {1}, volume = {8}, pages = {67--100}, year = {2005}, url = {http://www.kluweronline.com/issn/1386-4564}, abstract = {The intuition that different text classifiers behave in qualitatively different ways has long motivated attempts to build a better metaclassifier via some combination of classifiers. We introduce a probabilistic method for combining classifiers that considers the context-sensitive reliabilities of contributing classifiers. The method harnesses reliability indicators-variables that provide signals about the performance of classifiers in different situations. We provide background, present procedures for building metaclassifiers that take into consideration both reliability indicators and classifier outputs, and review a set of comparative studies undertaken to evaluate the methodology.}, } @inProceedings{Bickel04, author = {Steffen Bickel and Tobias Scheffer}, title = {Learning from message pairs for automatic email answering}, booktitle = {Proceedings of ECML-04, 15th European Conference on Machine Learning}, editor = {Jean-Fran{\c{c}}ois Boulicaut and Floriana Esposito and Fosca Giannotti and Dino Pedreschi}, address = {Pisa, IT}, pages = {87--98}, year = {2004}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 3201}, url = {}, abstract = {}, } @inProceedings{Biebricher88, author = {Peter Biebricher and Norbert Fuhr and Gerhard Knorz and Gerhard Lustig and Michael Schwantner}, title = {The automatic indexing system {AIR/PHYS}. From research to application}, booktitle = {Proceedings of SIGIR-88, 11th ACM International Conference on Research and Development in Information Retrieval}, editor = {Yves Chiaramella}, publisher = {ACM Press, New York, US}, address = {Grenoble, FR}, pages = {333--342}, year = {1988}, note = {Reprinted in Karen Sparck Jones and Peter Willett (eds.), ``Readings in Information Retrieval'', Morgan Kaufmann, San Francisco, US, 1997, pp.\ 513--517.}, url = {http://www.acm.org/pubs/articles/proceedings/ir/62437/p333-biebricher/p333-biebricher.pdf}, abstract = {Since October 1985, the automatic indexing system AIR/PHYS has been used in the input production of the physics data base of the Fachinformationszentrum Karlsruhe/West Germany. The texts to be indexed are abstracts written in English. The system of descriptors is prescribed. For the application of the AIR/PHYS system a large-scale dictionary containing more than 600000 word-descriptor relations resp. phrase-descriptor relations has been developed. Most of these relations have been obtained by means of statistical and heuristical methods. In consequence, the relation system is rather imperfect. Therefore, the indexing system needs some fault-tolerating features. An appropriate indexing approach and the corresponding structure of the AIR/PHYS system are described. Finally, the conditions of the application as well as problems of further development are discussed.}, } @inProceedings{Bigi03, author = {Brigitte Bigi}, title = {Using Kullback-Leibler distance for text categorization}, booktitle = {Proceedings of ECIR-03, 25th European Conference on Information Retrieval}, publisher = {Springer Verlag}, editor = {Fabrizio Sebastiani}, address = {Pisa, IT}, year = {2003}, pages = {305--319}, url = {http://link.springer.de/link/service/series/0558/papers/2633/26330305.pdf}, abstract = {A system that performs text categorization aims to assign appropriate categories from a predefined classification scheme to incoming documents. These assignments might be used for varied purposes such as filtering, or retrieval. This paper introduces a new effective model for text categorization with great corpus (more or less 1 million documents). Text categorization is performed using the Kullback-Leibler distance between the probability distribution of the document to classify and the probability distribution of each category. Using the same representation of categories, experiments show a significant improvement when the above mentioned method is used. KLD method achieve substantial improvements over the tfidf performing method.}, } @article{Blei03, author = {David M. Blei and Andrew Y. Ng and Michael I. Jordan}, title = {Latent Dirichlet Allocation}, journal = {Journal of Machine Learning Research}, volume = {3}, pages = {993--1022}, year = {2003}, url = {http://www.ai.mit.edu/projects/jmlr/papers/volume3/blei03a/blei03a.pdf}, abstract = {We describe latent Dirichlet allocation (LDA), a generative probabilistic model for collections of discrete data such as text corpora. LDA is a three-level hierarchical Bayesian model, in which each item of a collection is modeled as a finite mixture over an underlying set of topics. Each topic is, in turn, modeled as an infinite mixture over an underlying set of topic probabilities. In the context of text modeling, the topic probabilities provide an explicit representation of a document. We present efficient approximate inference techniques based on variational methods and an EM algorithm for empirical Bayes parameter estimation. We report results in document modeling, text classification, and collaborative filtering, comparing to a mixture of unigrams model and the probabilistic LSI model.}, } @article{Bloedorn98, author = {Eric Bloedorn and Ryszard S. Michalski}, title = {Data-Driven Constructive Induction}, journal = {IEEE Intelligent Systems}, year = {1998}, number = {2}, volume = {13}, pages = {30--37}, url = {}, abstract = {An inductive learning program's ability to find an accurate hypothesis can depend on the quality of the representation space. The authors developed a data-driven constructive-induction method that uses multiple operators to improve the representation space. They applied it to two real-world problems.}, } @inProceedings{Blosseville92, author = {M.J. Blosseville and Georges Hebrail and M.G. Montell and N. Penot}, title = {Automatic document classification: natural langage processing and expert system techniques used together}, booktitle = {Proceedings of SIGIR-92, 15th ACM International Conference on Research and Development in Information Retrieval}, editor = {Nicholas J. Belkin and Peter Ingwersen and Annelise Mark Pejtersen}, publisher = {ACM Press, New York, US}, address = {Kobenhavn, DK}, pages = {51--57}, year = {1992}, url = {http://www.acm.org/pubs/articles/proceedings/ir/133160/p51-blosseville/p51-blosseville.pdf}, abstract = {In this paper we describe an automated method of classifying research project descriptions: a human expert classifies a sample set of projects into a set of disjoint and pre-defined classes, and then the computer learns from this sample how to classify new projects into these classes. Both textual and non-textual information associated with the projects are used in the learning and classification phases. Textual information is processed by two methods of analysis: a natural language analysis followed by a statistical analysis. Non-textual information is processed by a symbolic learning technique. We present the results of some experiments done on real data: two different classifications of our research projects.}, } @article{Borko63, author = {Harold Borko and Myrna Bernick}, title = {Automatic document classification}, journal = {Journal of the Association for Computing Machinery}, year = {1963}, volume = {10}, number = {2}, pages = {151--161}, url = {http://www.acm.org/pubs/articles/journals/jacm/1963-10-2/p151-borko/p151-borko.pdf}, } @article{Borko64, author = {Harold Borko and Myrna Bernick}, title = {Automatic document classification. Part II: additional experiments}, journal = {Journal of the Association for Computing Machinery}, year = {1964}, volume = {11}, number = {2}, pages = {138--151}, url = {http://www.acm.org/pubs/articles/journals/jacm/1964-11-2/p138-borko/p138-borko.pdf}, abstract = {This study reports the results of a series of experiments in the techniques of automatic document classifications. Two different classification schedules are compared along with two methods of automatically classifying documents into categories. It is concluded that, while there is no significant difference in the predictive efficiency between the Bayesian and the Factor Score methods, automatic document classification is enhanced by the use of a factor-analytically-derived classification schedule. Approximately 55 percent of the documents were automatedly and correctly classified.}, } @inProceedings{Brank02a, author = {Janez Brank and Marko Grobelnik and Natasa Mili{\'{c}}-Frayling and Dunja Mladeni{\'{c}}}, title = {Feature selection using support vector machines}, booktitle = {Proceedings of the 3rd International Conference on Data Mining Methods and Databases for Engineering, Finance, and Other Fields}, year = {2002}, pages = {}, address = {Bologna, IT}, url = {http://www.brank.org/msr/FsNormal/Bologna/bologna-paper-4.pdf}, abstract = {Text categorization is the task of classifying natural language documents into a set of predefined categories. Documents are typically represented by sparse vectors under the vector space model, where each word in the vocabulary is mapped to one coordinate axis and its occurrence in the document gives rise to one nonzero component in the vector representing that document. When training classifiers on large collections of documents, both the time and memory requirements connected with processing of these vectors may be prohibitive. This calls for using a feature selection method, not only to reduce the number of features but also to increase the sparsity of document vectors. We propose a feature selection method based on linear Support Vector Machines (SVMs). First, we train the linear SVM on a subset of training data and retain only those features that correspond to highly weighted components (in absolute value sense) of the normal to the resulting hyperplane that separates positive and negative examples. This reduced feature space is then used to train a classifier over a larger training set because more documents now fit into the same amount of memory. In our experiments we compare the effectiveness of the SVM -based feature selection with that of more traditional feature selection methods, such as odds ratio and information gain, in achieving the desired tradeoff between the vector sparsity and the classification performance. Experimental results indicate that, at the same level of vector sparsity, feature selection based on SVM normals yields better classification performance than odds ratio- or information gainbased feature selection when linear SVM classifiers are used.}, } @inProceedings{Bruckner97, author = {T. Bruckner}, title = {The text categorization system TEKLIS at {TREC-6}}, booktitle = {Proceedings of TREC-6, 6th Text Retrieval Conference}, publisher = {National Institute of Standards and Technology, Gaithersburg, US}, editor = {Ellen M. Voorhees and Donna K. Harman}, year = {1997}, address = {Gaithersburg, US}, pages = {619--621}, url = {http://trec.nist.gov/pubs/trec6/papers/siemens.ps.gz}, abstract = {The article documents the author's participation in the filtering and routing tasks of TREC-6 with the commercial filtering system TEKLIS. TEKLIS is a training based statistical categorization system which incorporates shallow linguistic processing and fuzzy set methods. The author presents the core technology of TEKLIS, the results on the filtering and routing tasks and a discussion of the insights gained through participation in the exercise.}, } @inProceedings{Cai03, author = {Lijuan Cai and Thomas Hofmann}, title = {Text categorization by boosting automatically extracted concepts}, booktitle = {Proceedings of SIGIR-03, 26th ACM International Conference on Research and Development in Information Retrieval}, editor = {Jamie Callan and Gordon Cormack and Charles Clarke and David Hawking and Alan Smeaton}, publisher = {ACM Press, New York, US}, address = {Toronto, CA}, year = {2003}, pages = {182--189}, url = {http://doi.acm.org/10.1145/860435.860470}, abstract = {Term-based representations of documents have found wide-spread use in information retrieval. However, one of the main shortcomings of such methods is that they largely disregard lexical semantics and, as a consequence, are not sufficiently robust with respect to variations in word usage. In this paper we investigate the use of concept-based document representations to supplement word- or phrase-based features. The utilized concepts are automatically extracted from documents via probabilistic latent semantic analysis. We propose to use AdaBoost to optimally combine weak hypotheses based on both types of features. Experimental results on standard benchmarks confirm the validity of our approach, showing that AdaBoost achieves consistent improvements by including additional semantic features in the learned ensemble.}, } @inProceedings{Cai04, author = {Lijuan Cai and Thomas Hofmann}, title = {Hierarchical Document Categorization with Support Vector Machines}, booktitle = {Proceedings of CIKM-04, 13th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, address = {Washington, US}, editor = {David A. Evans and Luis Gravano and Otthein Herzog and ChengXiang Zhai and Marc Ronthaler}, year = {2004}, pages = {78--87}, url = {}, abstract = {}, } @inProceedings{Calado03, author = {P{\'{a}}vel Calado and Marco Cristo and Edleno Silva De Moura and Nivio Ziviani and Berthier A. Ribeiro-Neto and Marcos Andr{\'{e}} Gon{\c{c}}alves}, title = {Combining link-based and content-based methods for Web document classification}, booktitle = {Proceedings of CIKM-03, 12th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {2003}, address = {New Orleans, US}, pages = {394--401}, url = {http://doi.acm.org/10.1145/956863.956938}, abstract = {This paper studies how link information can be used to improve classification results for Web collections. We evaluate four different measures of subject similarity, derived from the Web link structure, and determine how accurate they are in predicting document categories. Using a Bayesian network model, we combine these measures with the results obtained by traditional content-based classifiers. Experiments on a Web directory show that best results are achieved when links from pages outside the directory are considered. Link information alone is able to obtain gains of up to 46 points in F1, when compared to a traditional content-based classifier. The combination with content-based methods can further improve the results, but too much noise may be introduced, since the text of Web pages is a much less reliable source of information. This work provides an important insight on which measures derived from links are more appropriate to compare Web documents and how these measures can be combined with content-based algorithms to improve the effectiveness of Web classification.}, } @inProceedings{Caldon03, author = {Patrick Caldon}, title = {Using Text Classification to Predict the Gene Knockout Behaviour of {S.\ Cerevisiae}}, booktitle = {Proceedings of APBC-03, 1st Asia-Pacific Bioinformatics Conference}, editor = {Yi-Ping P. Chen}, publisher = {Australian Computer Society}, address = {Adelaide, AU}, year = {2003}, pages = {211--214}, url = {}, abstract = {}, } @article{Carbonell00, author = {Jaime Carbonell and William W. Cohen and Yiming Yang}, title = {Guest editors' introduction to the special issue on machine learning and information retrieval}, journal = {Machine Learning}, volume = {39}, number = {2/3}, pages = {99--101}, year = {2000}, url = {http://www.wkap.nl/article.pdf?255754}, } @inProceedings{Cardoso03, author = {Ana Cardoso-Cachopo and Arlindo L. Oliveira}, title = {An Empirical Comparison of Text Categorization Methods}, booktitle = {Proceedings of SPIRE-03, 10th International Symposium on String Processing and Information Retrieval}, editor = {Mario A. Nascimento and Edleno S. De Moura and Arlindo L. Oliveira}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2857}, year = {2003}, address = {Manaus, BR}, pages = {183--196}, url = {http://www.gia.ist.utl.pt/~acardoso/spire03.pdf}, abstract = {In this paper we present a comprehensive comparison of the performance of a number of text categorization methods in two different data sets. In particular, we evaluate the Vector and Latent Semantic Analysis (LSA) methods, a classifier based on Support Vector Machines (SVM) and the k-Nearest Neighbor variations of the Vector and LSA models. We report the results obtained using the Mean Reciprocal Rank as a measure of overall performance, a commonly used evaluation measure for question answering tasks. We argue that this evaluation measure is also very well suited for text categorization tasks. Our results show that overall, SVMs and k-NN LSA perform better than the other methods, in a statistically significant way.}, } @inCollection{Caropreso01, author = {Maria Fernanda Caropreso and Stan Matwin and Fabrizio Sebastiani}, title = {A learner-independent evaluation of the usefulness of statistical phrases for automated text categorization}, year = {2001}, booktitle = {Text Databases and Document Management: Theory and Practice}, editor = {Amita G. Chin}, publisher = {Idea Group Publishing}, address = {Hershey, US}, pages = {78--102}, url = {http://www.math.unipd.it/~fabseb60/Publications/TD01a.pdf}, abstract = {In this work we investigate the usefulness of {\em $n$-grams} for document indexing in text categorization (TC). We call $n$-gram a set $g_k$ of $n$ word stems, and we say that $g_k$ occurs in a document $d_j$ when a sequence of words appears in $d_j$ that, after stop word removal and stemming, consists exactly of the $n$ stems in $g_k$, in some order. Previous researches have investigated the use of $n$-grams (or some variant of them) in the context of specific learning algorithms, and thus have not obtained general answers on their usefulness for TC. In this work we investigate the usefulness of $n$-grams in TC independently of any specific learning algorithm. We do so by applying feature selection to the pool of all $k$-grams ($k\leq n$), and checking how many $n$-grams score high enough to be selected in the top $\sigma$ $k$-grams. We report the results of our experiments, using various feature selection measures and varying values of $\sigma$, performed on the {\sc Reuters-21578} standard TC benchmark. We also report results of making actual use of the selected $n$-grams in the context of a linear classifier induced by means of the Rocchio method.}, } @inProceedings{Carreras01, author = {Xavier Carreras and Llu\'{\i}s M\'arquez}, title = {Boosting Trees for Anti-Spam Email Filtering}, year = {2001}, editor = {}, booktitle = {Proceedings of RANLP-01, 4th International Conference on Recent Advances in Natural Language Processing}, address = {Tzigov Chark, BG}, pages = {}, url = {http://www.lsi.upc.es/~carreras/pub/boospam.ps}, } @inProceedings{Cavnar94, author = {William B. Cavnar and John M. Trenkle}, title = {N-Gram-Based Text Categorization}, booktitle = {Proceedings of SDAIR-94, 3rd Annual Symposium on Document Analysis and Information Retrieval}, publisher = {}, editor = {}, year = {1994}, address = {Las Vegas, US}, pages = {161--175}, url = {http://www.nonlineardynamics.com/trenkle/papers/sdair-94-bc.ps.gz}, abstract = {Text categorization is a fundamental task in doc-ument processing, allowing the automated handling of enormous streams of documents in electronic form. One difficulty in handling some classes of documents is the presence of different kinds of textual errors, such as spelling and grammatical errors in email, and character recognition errors in documents that come through OCR. Text categorization must work reliably on all input, and thus must tolerate some level of these kinds of problems. We describe here an N-gram-based approach to text categorization that is tolerant of textual errors. The system is small, fast and robust. This system worked very well for language classification, achieving in one test a 99.8\% correct classification rate on Usenet newsgroup articles written in different languages. The system also worked reasonably well for classifying articles from a number of different computer-oriented newsgroups according to subject, achieving as high as an 80\% correct classification rate. There are also several obvious directions for improving the system's classification performance in those cases where it did not do as well. The system is based on calculating and comparing profiles of N-gram frequencies. First, we use the system to compute profiles on training set data that represent the various categories, e.g., language samples or newsgroup content samples. Then the system computes a profile for a particular document that is to be classified. Finally, the system computes a distance measure between the document's profile and each of the category profiles. The system selects the category whose profile has the smallest distance to the document's profile. The profiles involved are quite small, typically 10K bytes for a category training set, and less than 4K bytes for an individual document. Using N-gram frequency profiles provides a simple and reliable way to categorize documents in a wide range of classification tasks.}, } @inProceedings{Ceci03, author = {Michelangelo Ceci and Donato Malerba}, title = {Hierarchical Classification of HTML Documents with {WebClassII}}, booktitle = {Proceedings of ECIR-03, 25th European Conference on Information Retrieval}, publisher = {Springer Verlag}, editor = {Fabrizio Sebastiani}, address = {Pisa, IT}, year = {2003}, pages = {57--72}, url = {http://link.springer.de/link/service/series/0558/papers/2633/26330057.pdf}, abstract = {This paper describes a new method for the classification of a HTML document into a hierarchy of categories. The hierarchy of categories is involved in all phases of automated document classification, namely feature extraction, learning, and classification of a new document. The innovative aspects of this work are the feature selection process, the automated threshold determination for classification scores, and an experimental study on real-word Web documents that can be associated to any node in the hierarchy. Moreover, a new measure for the evaluation of system performances has been introduced in order to compare three different techniques (flat, hierarchical with proper training sets, hierarchical with hierarchical training sets). The method has been implemented in the context of a client-server application, named WebClassII. Results show that for hierarchical techniques it is better to use hierarchical training sets.}, } @inProceedings{Cerny83, author = {Barbara A. Cerny and Anna Okseniuk and J. Dennis Lawrence}, title = {A fuzzy measure of agreement between machine and manual assignment of documents to subject categories}, booktitle = {Proceedings of ASIS-83, 46th Annual Meeting of the American Society for Information Science}, publisher = {American Society for Information Science, Washington, US}, editor = {Raymond F. Vondran and Anne Caputo and Carol Wasserman and Richard A. Diener}, year = {1983}, address = {Washington, US}, pages = {265}, url = {}, } @inProceedings{Chai02, author = {Kian M. Chai and Hwee T. Ng and Hai L. Chieu}, title = {Bayesian online classifiers for text classification and filtering}, booktitle = {Proceedings of SIGIR-02, 25th ACM International Conference on Research and Development in Information Retrieval}, editor = {Micheline Beaulieu and Ricardo Baeza-Yates and Sung Hyon Myaeng and Kalervo J{\"{a}}rvelin}, publisher = {ACM Press, New York, US}, address = {Tampere, FI}, year = {2002}, pages = {97--104}, url = {http://doi.acm.org/10.1145/564376.564395}, abstract = {This paper explores the use of Bayesian online classifiers to classify text documents. Empirical results indicate that these classifiers are comparable with the best text classification systems. Furthermore, the online approach offers the advantage of continuous learning in the batch-adaptive text filtering task.}, } @inProceedings{Chakrabarti02, author = {Soumen Chakrabarti and Shourya Roy and Mahesh Soundalgekar}, title = {Fast and accurate text classification via multiple linear discriminant projections}, booktitle = {Proceedings of VLDB-02, 28th International Conference on Very Large Data Bases}, publisher = {}, editor = {}, year = {2002}, address = {Hong Kong, CN}, pages = {658--669}, url = {http://www.vldb.org/conf/2002/S19P01.pdf}, abstract = {Support vector machines (SVMs) have shown superb performance for text classification tasks. They are accurate, robust, and quick to apply to test instances. Their only potential drawback is their training time and memory requirement. For n training instances held in memory, the best-known SVM implementations take time proportional to n a , where a is typically between 1.8 and 2.1. SVMs have been trained on data sets with several thousand instances, but Web directories today contain millions of instances which are valuable for mapping billions of Web pages into Yahoo!-like directories. We present SIMPL, a nearly linear-time classification algorithm which mimics the strengths of SVMs while avoiding the training bottleneck. It uses Fisher's linear discriminant, a classical tool from statistical pattern recognition, to project training instances to a carefully selected low-dimensional subspace before inducing a decision tree on the projected instances. SIMPL uses efficient sequential scans and sorts, and is comparable in speed and memory scalability to widely-used naive Bayes (NB) classifiers, but it beats NB accuracy decisively. It not only approaches and sometimes exceeds SVM accuracy, but also beats SVM running time by orders of magnitude. While developing SIMPL, we also make a detailed experimental analysis of the cache performance of SVMs.}, } @inProceedings{Chakrabarti97, author = {Soumen Chakrabarti and Byron E. Dom and Rakesh Agrawal and Prabhakar Raghavan}, title = {Using taxonomy, discriminants, and signatures for navigating in text databases}, booktitle = {Proceedings of VLDB-97, 23rd International Conference on Very Large Data Bases}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, editor = {Matthias Jarke and Michael J. Carey and Klaus R. Dittrich and Frederick H. Lochovsky and Pericles Loucopoulos and Manfred A. Jeusfeld}, year = {1997}, address = {Athens, GR}, pages = {446--455}, url = {http://www.vldb.org/conf/1997/P446.PDF}, note = {An extended version appears as~\cite{Chakrabarti98c}}, abstract = {We explore how to organize a text database hierarchically to aid better searching and browsing. We propose to exploit the natural hierarchy of topics, or taxonomy, that many corpora, such as internet directories, digital libraries, and patent databases enjoy. In our system, the user navigates through the query response not as a flat unstructured list, but embedded in the familiar taxonomy, and annotated with document signatures computed dynamically with respect to where the user is located at any time. We show how to update such databases with new documents with high speed and accuracy. We use techniques from statistical pattern recognition to efficiently separate the feature words or discriminants from the noise words at each node of the taxonomy. Using these, we build a multi-level classifier. At each node, this classifier can ignore the large number of noise words in a document. Thus the classifier has a small model size and is very fast. However, owing to the use of context-sensitive features, the classifier is very accurate. We report on experiences with the Reuters newswire benchmark, the US Patent database, and web document samples from {{\sc Yahoo!}}\.}, } @inProceedings{Chakrabarti98b, author = {Soumen Chakrabarti and Byron E. Dom and Piotr Indyk}, title = {Enhanced hypertext categorization using hyperlinks}, booktitle = {Proceedings of SIGMOD-98, ACM International Conference on Management of Data}, editor = {Laura M. Haas and Ashutosh Tiwary}, publisher = {ACM Press, New York, US}, address = {Seattle, US}, year = {1998}, pages = {307--318}, url = {http://www.acm.org/pubs/articles/proceedings/mod/276304/p307-chakrabarti/p307-chakrabarti.pdf}, abstract = {A major challenge in indexing unstructured hypertext databases is to automatically extract meta-data that enables structured searching using topic taxonomies, circumvents keyword ambiguity and improves the quality of searching and profile-based routing and filtering. Therefore, an accurate classifier is an essential component of a hypertext database. Hyperlinks pose new problems not addressed in the extensive text classification literature. Links clearly contain high-quality semantic clues that are lost upon a purely term-based classifier, but exploiting link information is non-trivial because it is noisy. Naive use of terms in the link neighborhood of a document can even degrade accuracy. Our contribution is to propose robust statistical models and a relaxation labeling technique for better classification by exploiting link information in a small neighborhood around documents. Our technique also adapts gracefully to the fraction of neighboring documents having known topics. We experimented with pre-classified samples from {{\sc Yahoo!}}\ and the US Patent Database. We have developed a text classifier that misclassifies only 13\% of the documents in the Reuters benchmark; this is comparable to the best results ever obtained. Our new classifier misclassified 36\% of the patents, indicating that classifying hypertext can be more difficult than classifying text. Naively using terms in neighboring documents increased the error to 38\%; our hypertext classifier reduced it to 21\%. Results with the Yahoo! sample were more dramatic: the text classifier showed a 68\% error, whereas our hypertext classifier reduced this to just 21\%.}, } @article{Chakrabarti98c, author = {Soumen Chakrabarti and Byron E. Dom and Rakesh Agrawal and Prabhakar Raghavan}, title = {Scalable feature selection, classification and signature generation for organizing large text databases into hierarchical topic taxonomies}, journal = {Journal of Very Large Data Bases}, year = {1998}, number = {3}, volume = {7}, pages = {163--178}, url = {http://www.cs.berkeley.edu/~soumen/VLDB54_3.PDF}, abstract = {We explore how to organize large text databases hierarchically by topic to aid better searching, browsing and filtering. Many corpora, such as internet directories, digital libraries, and patent databases are manually organized into topic hierarchies, also called taxonomies. Similar to indices for relational data, taxonomies make search and access more efficient. However, the exponential growth in the volume of on-line textual information makes it nearly impossible to maintain such taxonomic organization for large, fast-changing corpora by hand. We describe an automatic system that starts with a small sample of the corpus in which topics have been assigned by hand, and then updates the database with new documents as the corpus grows, assigning topics to these new documents with high speed and accuracy. To do this, we use techniques from statistical pattern recognition to efficiently separate the feature words, or discriminants, from thenoise words at each node of the taxonomy. Using these, we build a multilevel classifier. At each node, this classifier can ignore the large number of ``noise'' words in a document. Thus, the classifier has a small model size and is very fast. Owing to the use of context-sensitive features, the classifier is very accurate. As a by-product, we can compute for each document a set of terms that occur significantly more often in it than in the classes to which it belongs. We describe the design and implementation of our system, stressing how to exploit standard, efficient relational operations like sorts and joins. We report on experiences with the Reuters newswire benchmark, the US patent database, and web document samples from Yahoo!. We discuss applications where our system can improve searching and filtering capabilities.}, } @article{Chakrabarti99, author = {Soumen Chakrabarti and Byron E. Dom and S. Ravi Kumar and Prabhakar Raghavan and Sridhar Rajagopalan and Andrew Tomkins and David Gibson and Jon Kleinberg}, title = {Mining the Web's link structure}, journal = {IEEE Computer}, year = {1999}, number = {8}, volume = {32}, pages = {60--67}, url = {http://dlib.computer.org/co/books/co1999/pdf/r8060.pdf}, abstract = {The Web is a hypertext body of approximately 300 million pages that continues to grow at roughly a million pages per day. Page variation is more prodigious than the data's raw scale: Taken as a whole, the set of Web pages lacks a unifying structure and shows far more authoring style and content variation than that seen in traditional text-document collections. This level of complexity makes an ``off-the-shelf'' database-management and information-retrieval solution impossible. To date, index-based search engines for the Web have been the primary tool by which users search for information. Such engines can build giant indices that let you quickly retrieve the set of all Web pages containing a given word or string. Experienced users can make effective use of such engines for tasks that can be solved by searching for tightly constrained keywords and phrases. These search engines are, however, unsuited for a wide range of equally important tasks. In particular, a topic of any breadth will typically contain several thousand or million relevant Web pages. How then, from this sea of pages, should a search engine select the correct ones-those of most value to the user?}, } @inProceedings{Chandrinos00, author = {Konstantinos V. Chandrinos and Ion Androutsopoulos and Georgios Paliouras and Constantine D. Spyropoulos}, title = {Automatic Web Rating: Filtering Obscene Content on the Web}, booktitle = {Proceedings of ECDL-00, 4th European Conference on Research and Advanced Technology for Digital Libraries}, editor = {Jos{\'e} L. Borbinha and Thomas Baker}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1923}, year = {2000}, address = {Lisbon, PT}, pages = {403--406}, url = {http://link.springer.de/link/service/series/0558/papers/1923/19230403.pdf}, abstract = {We present a method to detect automatically pornographic content on the Web. Our method combines techniques from language engineering and image analysis within a machine-learning framework. Experimental results show that it achieves nearly perfect performance on a set of hard cases.}, } @inProceedings{Chen00, author = {Hao Chen and Susan T. Dumais}, title = {Bringing order to the Web: automatically categorizing search results}, booktitle = {Proceedings of CHI-00, ACM International Conference on Human Factors in Computing Systems}, publisher = {ACM Press, New York, US}, editor = {}, year = {2000}, address = {Den Haag, NL}, pages = {145--152}, url = {http://www.acm.org/pubs/articles/proceedings/chi/332040/p145-chen/p145-chen.pdf}, abstract = {We developed a user interface that organizes Web search results into hierarchical categories. Text classification algorithms were used to automatically classify arbitrary search results into an existing category structure on-the-fly. A user study compared our new category interface with the typical ranked list interface of search results. The study showed that the category interface is superior both in objective and subjective measures. Subjects liked the category interface much better than the list interface, and they were 50\% faster at finding information that was organized into categories. Organizing search results allows users to focus on items in categories of interest rather than having to browse through all the results sequentially.}, } @inProceedings{Chen00a, author = {Hao Chen and Tin Kam Ho}, title = {Evaluation of Decision Forests on Text Categorization}, booktitle = {Proceedings of the 7th SPIE Conference on Document Recognition and Retrieval}, publisher = {SPIE {}-{} The International Society for Optical Engineering}, editor = {Daniel P. Lopresti and Jiangying Zhou}, year = {2000}, address = {San Jose, US}, pages = {191--199}, url = {http://cm.bell-labs.com/who/tkh/papers/textcat.ps.gz}, abstract = {Text categorization is useful for indexing documents for information retrieval, filtering parts for document understanding, and summarizing contents of documents of special interests. We describe a text categorization task and an experiment using documents from the Reuters and OHSUMED collections. We applied the Decision Forest classifier and compared its accuracies to those of C4.5 and kNN classifiers, using both category dependent and category independent term selection schemes. It is found that Decision Forest outperforms both C4.5 and kNN in all cases, and that category dependent term selection yields better accuracies. Performances of all three classifiers degrade from the Reuters collection to the OHSUMED collection, but Decision Forest remains to be superior.}, } @inProceedings{Chen01, author = {Chien Chin Chen and Chang Chen, Meng and Yeali Sun}, title = {PVA: A Self-Adaptive Personal View Agent}, booktitle = {Proceedings of KDD-01, 7th ACM SIGKDD International Conferece on Knowledge Discovery and Data Mining}, editor = {Foster Provost and Ramakrishnan Srikant}, year = {2001}, pages = {257--262}, publisher = {ACM Press, New York, US}, address = {San Francisco, US}, url = {http://doi.acm.org/10.1145/502512.502548}, abstract = {In this paper, we present PVA, an adaptive personal view information agent system to track, learn and manage, user's interests in Internet documents. When user's interests change, PVA, in not only the contents, but also in the structure of user profile, is modified to adapt to the changes. Experimental results show that modulating the structure of user profile does increase the accuracy of personalization systems.}, } @article{Chen02, author = {Chien Chin Chen and Chang Chen, Meng and Yeali Sun}, title = {PVA: A Self-Adaptive Personal View Agent}, journal = {Journal of Intelligent Information Systems}, year = {2002}, note = {Special Issue on Automated Text Categorization}, volume = {18}, number = {2/3}, pages = {173--194}, url = {http://www.wkap.nl/article.pdf?391245}, abstract = {In this paper, we present PVA, an adaptive personal view information agent system for tracking, learning and managing user interests in Internet documents. PVA consists of three parts: a {\it proxy}, {\it personal view constructor}, and {\it personal view maintainer}. The proxy logs the user's activities and extracts the user's interests without user intervention. The personal view constructor mines user interests and maps them to a class hierarchy (i.e., personal view). The personal view maintainer synchronizes user interests and the personal view periodically. When user interests change, in PVA, not only the contents, but also the structure of the user profile are modified to adapt to the changes. In addition, PVA considers the aging problem of user interests. The experimental results show that modulating the structure of the user profile increases the accuracy of a personalization system.}, } @article{Chen03, author = {Chen L. and Tokuda N. and Nagai A.}, title = {A new differential LSI space-based probabilistic document classifier}, journal = {Information Processing Letters}, pages = {203--212}, year = {2003}, volume = {88}, number = {5}, doi = {http://dx.doi.org/10.1016/j.ipl.2003.09.002}, abstract = {We have developed a new effective probabilistic classifier for document classification by introducing the concept of differential document vectors and DLSI (differential latent semantic indexing) spaces. A combined use of the projections on and the distances to the DLSI spaces introduced from the differential document vectors improves the adaptability of the LSI (latent semantic indexing) method by capturing unique characteristics of documents. Using the intra- and extra-document statistics, both a simple posteriori calculation on a small example and an experiment on a large Reuters-21578 database demonstrate the advantage of the DLSI space-based probabilistic classifier over the LSI space-based classifier in classification performance.}, } @inProceedings{Chen04, author = {Wenliang Chen and Jingbo Zhu and Honglin Wu and Yao Tianshun}, title = {Automatic Learning Features Using Bootstrapping for Text Categorization}, booktitle = {Proceedings of CICLING-04, 5th International Conference on Computational Linguistics and Intelligent Text Processing}, year = {2004}, editor = {Alexander F. Gelbukh}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Seoul, KO}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2945}, pages = {571--579}, url = {}, abstract = {}, } @inProceedings{Cheng01, author = {Cheng, Chun-Hung and Jian Tang and Ada Wai-Chee and Irwin King}, title = {Hierarchical Classification of Documents with Error Control}, booktitle = {Proceedings of PAKDD-01, 5th Pacific-Asia Conferenece on Knowledge Discovery and Data Mining}, editor = {David Cheung and Qing Li and Graham Williams}, year = {2001}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Hong Kong, CN}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2035}, pages = {433--443}, url = {http://link.springer-ny.com/link/service/series/0558/papers/2035/20350433.pdf}, abstract = {Classification is a function that matches a new object with one of the predefined classes. Document classification is characterized by the large number of attributes involved in the objects (documents). The traditional method of building a single classifier to do all the classification work would incur a high overhead. Hierarchical classifi- cation is a more efficient method - instead of a single classifier, we use a set of classifiers distributed over a class taxonomy, one for each internal node. However, once a misclassification occurs at a high level class, it may result in a class that is far apart from the correct one. An existing approach to coping with this problem requires terms also to be arranged hierarchically. In this paper, instead of overhauling the classifier itself, we propose mechanisms to detect misclassification and take appropriate actions. We then discuss an alternative that masks the misclassification based on a well known software fault tolerance technique. Our experiments show our algorithms represent a good trade-off between speed and accuracy in most applications.}, } @inProceedings{Cheong02, author = {Cheong Fung, Gabriel P. and Jeffrey X. Yu and Hongjun Lu}, title = {Discriminative Category Matching: Efficient Text Classification for Huge Document Collections}, booktitle = {Proceedings of ICDM-02, 2nd IEEE International Conference on Data Mining}, editor = {}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, address = {Maebashi City, JP}, year = {2002}, pages = {187--194}, url = {http://dlib.computer.org/conferen/icdm/1754/pdf/17540187.pdf}, abstract = {With the rapid growth of textual information available on the Internet, having a good model for classifying and managing documents automatically is undoubtly important. When more documents are archived, new terms, new concepts and concept-drift will frequently appear. Without a doubt, updating the classification model frequently rather than using the old model for a very long period is absolutely essential. Here, the challenges are: a) obtain a high accuracy classification model; b) consume low computational time for both model training and operation; and c) occupy low storage space. However, none of the existing classification approaches could achieve all of these requirements. In this paper, we propose a novel text classification approach, called Discriminative Category Matching, which could achieve all of the stated characteristics. Extensive experiments using two benchmarks and a large real-life collection are conducted. The encouraging results indicated that our approach is hignhly feasible.}, } @article{Chouchoulas01, author = {Alexios Chouchoulas and Qiang Shen}, title = {Rough set-aided keyword reduction for text categorization}, journal = {Applied Artificial Intelligence}, pages = {843--873}, year = {2001}, volume = {15}, number = {9}, url = {}, abstract = {The volume of electronically stored information increases exponentially as the state of the art progresses. Automated information filtering (IF) and information retrieval (IR) systems are therefore acquiring rapidly increasing prominence. However, such systems sacrifice efficiency to boost effectiveness. Such systems typically have to cope with sets of rectors of many tens of thousands of dimensions. Rough set (RS) theory can be applied to reducing the dimensionality of data used in IF/IR tasks, by providing a measure of the information content of datasets with respect to a given classification. This can aid IF/IR systems that rely on the acquisition of large numbers of term weights or other measures of relevance. This article investigates the applicability of RS theory to the IF/IR application domain and compares this applicability with respect to various existing TC techniques. The ability, of the approach to generalize, given a minimum of training data is also addressed. The background of RS theory is presented, with an illustrative example to demonstrate the operation of the RS-based dimensionality reduction. A modular system is proposed which allows the integration of this technique with a large variety of different IF/IR approaches. The example application, categorization of E-mail messages, is described. Systematic experiments and their results are reported and analyzed.}, } @inProceedings{Chuang00, author = {Wesley T. Chuang and Asok Tiyyagura and Jihoon Yang and Giovanni Giuffrida}, title = {A Fast Algorithm for Hierarchical Text Classification}, booktitle = {Proceedings of DaWaK-00, 2nd International Conference on Data Warehousing and Knowledge Discovery}, editor = {Yahiko Kambayashi and Mukesh Mohania and A.Min Tjoa}, year = {2000}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1874}, address = {London, UK}, pages = {409--418}, url = {http://www.cs.iastate.edu/~yang/Papers/dawak00.ps}, abstract = {Text classification is becoming more important with the proliferation of the Internet and the huge amount of data it transfers. We present an efficient algorithm for text classification using hierarchical classifiers based on a concept hierarchy. The simple TFIDF classifier is chosen to train sample data and to classify other new data. Despite its simplicity, results of experiments on Web pages and TV closed captions demonstrate high classification accuracy. Application of feature subset selection techniques improves the performance. Our algorithm is computationally efficient being bounded by O(n log n) forn samples.}, } @inProceedings{Ciravegna99, author = {Fabio Ciravegna and Alberto Lavelli and Nadia Mana and Johannes Matiasek and Luca Gilardoni and Silvia Mazza and William J. Black and Fabio Rinaldi}, title = {FACILE: Classifying Texts Integrating Pattern Matching and Information Extraction}, booktitle = {Proceedings of IJCAI-99, 16th International Joint Conference on Artificial Intelligence}, editor = {Thomas Dean}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, year = {1999}, pages = {890--895}, address = {Stockholm, SE}, url = {http://ecate.itc.it:1024/lavelli/lavelli-papers/IJCAI99/ijcai99.ps.gz}, abstract = {Successfully managing information means being able to find relevant new information and to correctly integrate it with pre-existing knowledge. Much information is nowadays stored as multilingual textual data; therefore advanced classification systems are currently considered as strategic components for effective knowledge management. We describe an experience integrating different innovative AI technologies such as hierarchical pattern matching and information extraction to provide flexible multilingual classification adaptable to user needs. Pattern matching produces fairly accurate and fast categorisation over a large number of classes, while information extraction provides fine-grained classification for a reduced number of classes. The resulting system was adopted by the main Italian financial news agency providing a pay-to-view service.}, } @inProceedings{Clack97, author = {Chris Clack and Johnny Farringdon and Peter Lidwell and Tina Yu}, title = {Autonomous document classification for business}, editor = {W. Lewis Johnson}, publisher = {ACM Press, New York, US}, booktitle = {Proceedings of the 1st International Conference on Autonomous Agents}, address = {Marina Del Rey, US}, year = {1997}, pages = {201--208}, url = {http://www.acm.org/pubs/articles/proceedings/ai/267658/p201-clack/p201-clack.pdf}, abstract = {With the continuing exponential growth of the Internet and the more recent growth of business Intranets, the commercial world is becoming increasingly aware of the problem of electronic information overload. This has encouraged interest in developing agents/softbots that can act as electronic personal assistants and can develop and adapt representations of users information needs, commonly known as profiles. As the result of collaborative research with Friends of the Earth, an environmental issues campaigning organisation, we have developed a general purpose information classification agent architecture and have applied it to the problem of document classification and routing. Collaboration with Friends of the Earth allows us to test our ideas in a non-academic context involving high volumes of documents. We use the technique of genetic programming (GP), (Koza and Rice 1992), to evolve classifying agents. This is a novel approach for document classification, where each agent evolves a parse-tree representation of a user's particular information need. The other unusual features of our research are the longevity of our agents and the fact that they undergo a continual training process; feedback from the user enables the agent to adapt to the user's long-term information requirements.}, } @inProceedings{Cohen95, author = {William W. Cohen}, title = {Text categorization and relational learning}, booktitle = {Proceedings of ICML-95, 12th International Conference on Machine Learning}, editor = {Armand Prieditis and Stuart J. Russell}, address = {Lake Tahoe, US}, year = {1995}, pages = {124--132}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.research.whizbang.com/~wcohen/postscript/ml-95-ir.ps}, abstract = {We evaluate the first order learning system FOIL on a series of text categorization problems. It is shown that FOIL usually forms classifiers with lower error rates and higher rates of precision and recall with a relational encoding than with a propositional encoding. We show that FOIL's performance can be improved by relation selection, a first order analog of feature selection. Relation selection improves FOIL's performance as measured by any of recall, precision, F-measure, or error rate. With an appropriate level of relation selection, FOIL appears to be competitive with or superior to existing propositional techniques.}, } @inCollection{Cohen95a, author = {William W. Cohen}, title = {Learning to classify English text with ILP methods}, booktitle = {Advances in inductive logic programming}, editor = {De Raedt, Luc}, publisher = {IOS Press}, address = {Amsterdam, NL}, pages = {124--143}, year = {1995}, url = {http://www.research.whizbang.com/~wcohen/postscript/ilp.ps}, abstract = {Text categorization is the task of classifying text into one of several predefined categories. In this paper we will evaluate the effectiveness of several ILP methods for text categorization, and also compare them to their propositional analogs. The methods considered are FOIL, the propositional rule-learning system RIPPER, and a first-order version of RIPPER called FLIPPER. We show that the benefit of using a first-order representation in this domain is relatively modest; in particular, the performance difference between FLIPPER and FOIL and their propositional counterparts is quite small, compared to the differences between FOIL and FLIPPER. However, a first-order representation seems to be advantageous when high-precision classifiers are desirable.}, } @inProceedings{Cohen96a, author = {William W. Cohen and Yoram Singer}, title = {Context-sensitive learning methods for text categorization}, booktitle = {Proceedings of SIGIR-96, 19th ACM International Conference on Research and Development in Information Retrieval}, editor = {Hans-Peter Frei and Donna Harman and Peter Sch{\"{a}}uble and Ross Wilkinson}, publisher = {ACM Press, New York, US}, year = {1996}, address = {Z{\"{u}}rich, CH}, pages = {307--315}, note = {An extended version appears as~\cite{Cohen99}}, url = {http://www.research.whizbang.com/~wcohen/postscript/sigir-96.ps}, abstract = {Two machine learning algorithms, RIPPER and sleeping experts for phrases, are evaluated on a number of large text categorization problems. These algorithms both construct classifiers that allow the ``context'' of a word w to affect how (or even whether) the presence or absence of w will contribute to a classification. However, RIPPER and sleeping experts differ radically in many other respects. Differences include: different notions as to what constitutes a context; different ways of combining contexts to construct a classifier; different methods to search for a combination of contexts; and different criteria as to what contexts should be included in such a combination. In spite of these differences, both RIPPER and sleeping experts perform extremely well across a wide variety of categorization problems, generally outperforming previously applied learning methods. We view this result as a confirmation of the usefulness of classifiers that represent contextual information.}, } @inProceedings{Cohen98, author = {William W. Cohen and Haym Hirsh}, title = {Joins that generalize: text classification using {{\sc Whirl}}}, booktitle = {Proceedings of KDD-98, 4th International Conference on Knowledge Discovery and Data Mining}, editor = {Rakesh Agrawal and Paul E. Stolorz and Gregory Piatetsky-Shapiro}, publisher = {AAAI Press, Menlo Park, US}, year = {1998}, address = {New York, US}, pages = {169--173}, url = {http://www.research.whizbang.com/~wcohen/postscript/kdd-98.ps}, abstract = {WHIRL is an extension of relational databases that can perform ``soft joins'' based on the similarity of textual identifiers; these soft joins extend the traditional operation of joining tables based on the equivalence of atomic values. This paper evaluates WHIRL on a number of inductive classification tasks using data from the World Wide Web. We show that although WHIRL is designed for more general similarity-based reasoning tasks, it is competitive with mature inductive classification systems on these classification tasks. In particular, WHIRL generally achieves lower generalization error than C4.5, RIPPER, and several nearest-neighbor methods. WHIRL is also fast-up to 500 times faster than C4.5 on some benchmark problems. We also show that WHIRL can be efficiently used to select from a large pool of unlabeled items those that can be classified correctly with high confidence.}, } @article{Cohen99, author = {William W. Cohen and Yoram Singer}, title = {Context-sensitive learning methods for text categorization}, journal = {ACM Transactions on Information Systems}, year = {1999}, volume = {17}, number = {2}, pages = {141--173}, url = {http://www.acm.org/pubs/articles/journals/tois/1999-17-2/p141-cohen/p141-cohen.pdf}, abstract = {Two recently implemented machine-learning algorithms, RIPPER and sleeping-experts for phrases, are evaluated on a number of large text categorization problems. These algorithms both construct classifiers that allow the ``context'' of a word w to affect how (or even whether) the presence or absence of w will contribute to a classification. However, RIPPER and sleeping-experts differ radically in many other respects: differences include different notions as to what constitutes a context, different ways of combining contexts to construct a classifier, different methods to search for a combination of contexts, and different criteria as to what contexts should be included in such a combination. In spite of these differences, both RIPPER and sleeping-experts perform extremely well across a wide variety of categorization problems, generally outperforming previously applied learning methods. We view this result as a confirmation of the usefulness of classifiers that represent contextual information.}, } @inProceedings{Crammer02, author = {Koby Crammer and Yoram Singer}, title = {A New Family of Online Algorithms for Category Ranking}, booktitle = {Proceedings of SIGIR-02, 25th ACM International Conference on Research and Development in Information Retrieval}, editor = {Micheline Beaulieu and Ricardo Baeza-Yates and Sung Hyon Myaeng and Kalervo J{\"{a}}rvelin}, publisher = {ACM Press, New York, US}, address = {Tampere, FI}, year = {2002}, pages = {151--158}, url = {http://doi.acm.org/10.1145/564376.564404}, abstract = {We describe a new family of topic-ranking algorithms for multi-labeled documents. The motivation for the algorithms stems from recent advances in online learning algorithms. The algorithms we present are simple to implement and are time and memory efficient. We evaluate the algorithms on the Reuters-21578 corpus and the new corpus released by Reuters in 2000. On both corpora the algorithms we present outperform adaptations to topic-ranking of Rocchio's algorithm and the Perceptron algorithm. We also outline the formal analysis of the algorithm in the mistake bound model. To our knowledge, this work is the first to report performance results with the entire new Reuters corpus.}, } @article{Craven00, author = {Mark Craven and Dan DiPasquo and Dayne Freitag and Andrew K. McCallum and Tom M. Mitchell and Kamal Nigam and Se{\'{a}}n Slattery}, title = {Learning to Construct Knowledge Bases from the World Wide Web}, journal = {Artificial Intelligence}, volume = {118}, number = {1/2}, year = {2000}, pages = {69--113}, url = {http://www.cs.cmu.edu/afs/cs.cmu.edu/project/theo-11/www/wwkb/overview-aij99.ps.gz}, abstract = {The World Wide Web is a vast source of information accessible to computers, but understandable only to humans. The goal of the research described here is to automatically create a computer understandable knowledge base whose content mirrors that of the World Wide Web. Such a knowledge base would enable much more effective retrieval of Web information, and promote new uses of the Web to support knowledge-based inference and problem solving. Our approach is to develop a trainable information extraction system that takes two inputs. The first is an ontology that defines the classes (e.g., company, person, employee, product) and relations (e.g., employed_by, produced_by) of interest when creating the knowledge base. The second is a set of training data consisting of labeled regions of hypertext that represent instances of these classes and relations. Given these inputs, the system learns to extract information from other pages and hyperlinks on the Web. This article describes our general approach, several machine learning algorithms for this task, and promising initial results with a prototype system that has created a knowledge base describing university people, courses, and research projects.}, } @article{Craven01, author = {Craven, Mark and Slattery, Se{\'{a}}n}, title = {Relational learning with statistical predicate invention: Better models for hypertext}, journal = {Machine Learning}, pages = {97--119}, year = {2001}, volume = {43}, number = {1/2}, url = {http://www.wkap.nl/article.pdf?321079}, abstract = {We present a new approach to learning hypertext classifiers that combines a statistical text-learning method with a relational rule learner. This approach is well suited to learning in hypertext domains because its statistical component allows it to characterize text in terms of word frequencies, whereas its relational component is able to describe how neighboring documents are related to each other by hyperlinks that connect them. We evaluate our approach by applying it to tasks that involve learning definitions for (i) classes of pages, (ii) particular relations that exist between pairs of pages, and (iii) locating a particular class of information in the internal structure of pages. Our experiments demonstrate that this new approach is able to learn more accurate classifiers than either of its constituent methods alone.}, } @inProceedings{Craven98, author = {Mark Craven and Dan DiPasquo and Dayne Freitag and Andrew K. McCallum and Tom M. Mitchell and Kamal Nigam and Se{\'{a}}n Slattery}, title = {Learning to extract symbolic knowledge from the World Wide Web}, booktitle = {Proceedings of AAAI-98, 15th Conference of the American Association for Artificial Intelligence}, publisher = {AAAI Press, Menlo Park, US}, year = {1998}, pages = {509--516}, address = {Madison, US}, note = {An extended version appears as~\cite{Craven00}}, url = {http://www.cs.cmu.edu/afs/cs.cmu.edu/project/theo-11/www/wwkb/overview-aaai98.ps.gz}, abstract = {The World Wide Web is a vast source of information accessible to computers, but understandable only to humans. The goal of the research described here is to automatically create a computer understandable world wide knowledge base whose content mirrors that of the World Wide Web. Such a knowledge base would enable much more effective retrieval of Web information, and promote new uses of the Web to support knowledge-based inference and problem solving. Our approach is to develop a trainable information extraction system that takes two inputs: an ontology defining the classes and relations of interest, and a set of training data consisting of labeled regions of hypertext representing instances of these classes and relations. Given these inputs, the system learns to extract information from other pages and hyperlinks on the Web. This paper describes our general approach, several machine learning algorithms for this task, and promising initial results with a prototype system.}, } @article{Creecy92, author = {Robert M. Creecy and Brij M. Masand and Stephen J. Smith and David L. Waltz}, title = {Trading MIPS and memory for knowledge engineering: classifying census returns on the Connection Machine}, journal = {Communications of the ACM}, volume = {35}, number = {8}, year = {1992}, pages = {48--63}, url = {http://www.acm.org/pubs/articles/journals/cacm/1992-35-8/p48-creecy/p48-creecy.pdf}, } @inProceedings{Cristianini01, author = {Nello Cristianini and John Shawe-Taylor and Huma Lodhi}, title = {Latent Semantic Kernels}, booktitle = {Proceedings of ICML-01, 18th International Conference on Machine Learning}, editor = {Carla Brodley and Andrea Danyluk}, address = {Williams College, US}, year = {2001}, pages = {66--73}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {}, abstract = {Kernel methods like Support Vector Machines have successfully been used for text categorization. A standard choice of kernel function has been the inner product between the vector-space representation of two documents, in analogy with classical information retrieval (IR) approaches. Latent Semantic Indexing (LSI) has been successfully used for IR purposes, as a technique for capturing semantic relations between terms and inserting them into the similarity measure between two documents. One of its main drawbacks, in IR, is its computational cost. In this paper we describe how the LSI approach can be implemented in a kernel-defined feature space. We provide experimental results demonstrating that the approach can significantly improve performance, and that it does not impair it.} } @inCollection{Cristianini01a, author = {Huma Lodhi and John Shawe-Taylor and Nello Cristianini and Christopher J. Watkins}, title = {Discrete Kernels for Text Categorisation}, booktitle = {Advances in Neural Information Processing Systems}, editor = {Todd K. Leen and Thomas G. Dietterich and Volker Tresp}, volume = {13}, year = {2001}, pages = {563--569}, publisher = {MIT Press, Cambridge, MA}, url = {http://www.support-vector.net/papers/LodhiShawe-TaylorCristianiniWatkins_ps.ps}, abstract = {}, } @article{Cristianini02, author = {Nello Cristianini and John Shawe-Taylor and Huma Lodhi}, title = {Latent Semantic Kernels}, journal = {Journal of Intelligent Information Systems}, year = {2002}, note = {Special Issue on Automated Text Categorization}, volume = {18}, number = {2/3}, pages = {127--152}, url = {http://www.wkap.nl/article.pdf?391243}, abstract = {Kernel methods like Support Vector Machines have successfully been used for text categorization. A standard choice of kernel function has been the inner product between the vector-space representation of two documents, in analogy with classical information retrieval (IR) approaches. Latent Semantic Indexing (LSI) has been successfully used for IR purposes as a technique for capturing semantic relations between terms and inserting them into the similarity measure between two documents. One of its main drawbacks, in IR, is its computational cost. In this paper we describe how the LSI approach can be implemented in a kernel-defined feature space. We provide experimental results demonstrating that the approach can significantly improve performance, and that it does not impair it.}, } @inProceedings{Dagan96, author = {Dagan, Ido and Feldman, Ronen and Hirsh, Haym}, title = {Keyword-based browsing and analysis of large document sets}, booktitle = {Proceedings of SDAIR-96, 5th Annual Symposium on Document Analysis and Information Retrieval}, publisher = {}, editor = {}, year = {1996}, address = {Las Vegas, US}, pages = {191--207}, url = {}, abstract = {Knowledge discovery in databases (KDD) focuses on the computerized exploration of large amounts of data and on the discovery of interesting patterns within them. While most work on KDD has been concerned with structured databases, there has been little work on handling the huge amount of information that is available only in unstructured textual form. The paper describes the KDT system for knowledge discovery in texts. It is built on top of a text-categorization paradigm where text articles are annotated with keywords organized in a hierarchical structure. Knowledge discovery is performed by analyzing the co-occurrence frequencies of keywords from this hierarchy in the various documents. The authors show how this term-frequency approach supports a range of KDD operations, providing a general framework for knowledge discovery and exploration in collections of unstructured text.}, } @inProceedings{Dagan97, author = {Ido Dagan and Yael Karov and Dan Roth}, title = {Mistake-driven learning in text categorization}, booktitle = {Proceedings of EMNLP-97, 2nd Conference on Empirical Methods in Natural Language Processing}, publisher = {Association for Computational Linguistics, Morristown, US}, editor = {Claire Cardie and Ralph Weischedel}, year = {1997}, address = {Providence, US}, pages = {55--63}, url = {http://l2r.cs.uiuc.edu/~danr/Papers/categ.ps.gz}, abstract = {Learning problems in the text processing domain often map the text to a space whose dimensions are the measured features of the text, e.g., its words. Three characteristic properties of this domain are (a) very high dimensionality, (b) both the learned concepts and the instances reside very sparsely in the feature space, and (c) a high variation in the number of active features in an instance. In this work we study three mistake-driven learning algorithms for a typical task of this nature - text categorization. We argue that these algorithms which categorize documents by learning a linear separator in the feature space have a few properties that make them ideal for this domain. We then show that a quantum leap in performance is achieved when we further modify the algorithms to better address some of the specific characteristics of the domain. In particular, we demonstrate (1) how variation in document length can be tolerated by either normalizing feature weights or by using negative weights, (2) the positive effect of applying a threshold range in training, (3) alternatives in considering feature frequency, and (4) the benefits of discarding features while training. Overall, we present an algorithm, a variation of Littlestone's Winnow, which performs significantly better than any other algorithm tested on this task using a similar feature set.}, } @inProceedings{Dalessio00, author = {Stephen D'Alessio and Keitha Murray and Robert Schiaffino and Aaron Kershenbaum}, title = {The effect of using Hierarchical classifiers in Text Categorization}, booktitle = {Proceeding of RIAO-00, 6th International Conference ``Recherche d'Information Assistee par Ordinateur''}, editor = {}, address = {Paris, FR}, year = {2000}, pages = {302--313}, url = {http://www.iona.edu/cs/FacultyPublications/riao2000New.pdf}, abstract = {Given a set of categories, with or without a preexisting hierarchy among them, we consider the problem of assigning documents to one or more of these categories from the point of view of a hierarchy with more or less depth. We can choose to make use of none, part or all of the hierarchical structure to improve the categorization effectiveness and efficiency. It is possible to create additional hierarchy among the categories. We describe a procedure for generating a hierarchy of classifiers that model the hierarchy structure. We report on computational experience using this procedure. We show that judicious use of a hierarchy can significantly improve both the speed and effectiveness of the categorization process. Using the Reuters-21578 corpus, we obtain an improvement in running time of over a factor of three and a 5\% improvement in F-measure.}, } @inProceedings{Dalessio98, author = {Stephen D'Alessio and Keitha Murray and Robert Schiaffino and Aaron Kershenbaum}, title = {Category Levels in Hierarchical Text Categorization}, booktitle = {Proceedings of EMNLP-98, 3rd Conference on Empirical Methods in Natural Language Processing}, year = {1998}, publisher = {Association for Computational Linguistics, Morristown, US}, editor = {}, pages = {}, address = {Granada, ES}, url = {http://www.iona.edu/cs/FacultyPublications/emnlpf.pdf}, abstract = {We consider the problem of assigning level numbers (weights) to hierarchically organized categories during the process of text categorization. These levels control the ability of the categories to attract documents during the categorization process. The levels are adjusted in order to obtain a balance between recall and precision for each category. If a category's recall exceeds its precision, the category is too strong and its level is reduced. Conversely, a category's level is increased to strengthen it if its precision exceeds its recall. The categorization algorithm used is a supervised learning procedure that uses a linear classifier based on the category levels. We are given a set of categories, organized hierarchically. We are also given a training corpus of documents already placed in one or more categories. From these, we extract vocabulary, words that appear with high frequency within a given category, characterizing each subject area. Each node's vocabulary is filtered and its words assigned weights with respect to the specific category. Then, test documents are scanned and categories ranked based on the presence of vocabulary terms. Documents are assigned to categories based on these rankings. We demonstrate that precision and recall can be significantly improved by solving the categorization problem taking hierarchy into account. Specifically, we show that by adjusting the category levels in a principled way, that precision can be significantly improved, from 84\% to 91\%, on the much-studied Reuters-21578 corpus organized in a three-level hierarchy of categories.}, } @article{Damashek95, author = {Marc Damashek}, title = {Gauging Similarity with N-Grams: Language-Independent Categorization of Text}, journal = {Science}, year = {1995}, volume = {267}, number = {5199}, pages = {843--848}, url = {}, abstract = {A language-independent means of gauging topical similarity in unrestricted text is described. The method combines information derived from n-grams (consecutive sequences of n characters) with a simple vector-space technique that makes sorting, categorization, and retrieval feasible in a large multilingual collection of documents. No prior information about document content or language is required. Context, as it applies to document similarity, can be accommodated by a well-defined procedure. When an existing document is used as an exemplar, the completeness and accuracy with which topically related documents are retrieved is comparable to that of the best existing systems. The results of a formal evaluation are discussed, and examples are given using documents in English and Japanese.}, } @article{Damerau04, author = {Fred J. Damerau and Tong Zhang and Sholom M. Weiss and Nitin Indurkhya}, title = {Text categorization for a comprehensive time-dependent benchmark}, journal = {Information Processing and Management}, year = {2004}, volume = {40}, number = {2}, pages = {209--221}, url = {}, abstract = {}, } @article{Dasigi01, author = {Dasigi, Venu and Mann, Reinhold C. and Protopopescu, Vladimir A.}, title = {Information fusion for text classification: an experimental comparison}, journal = {Pattern Recognition}, year = {2001}, volume = {34}, number = {12}, pages = {2413--2425}, url = {}, abstract = {This article reports on our experiments and results on the effectiveness of different feature sets and information fusion from some combinations of them in classifying free text documents into a given number of categories. We use different feature sets and integrate neural network learning into the method. The feature sets are based on the ``latent semantics'' of a reference library - a collection of documents adequately representing the desired concepts. We found that a larger reference library is not necessarily better. Information fusion almost always gives better results than the individual constituent feature sets, with certain combinations doing better than the others.}, } @inProceedings{Davidov04, author = {Dmitry Davidov and Evgeniy Gabrilovich and Shaul Markovitch}, title = {Parameterized Generation of Labeled Datasets for Text Categorization Based on a Hierarchical Directory}, booktitle = {Proceedings of SIGIR-04, 27th ACM International Conference on Research and Development in Information Retrieval}, editor = {Kalervo J{\"{a}}rvelin and James Allan and Peter Bruza and Mark Sanderson}, publisher = {ACM Press, New York, US}, address = {Sheffield, UK}, year = {2004}, pages = {250--257}, url = {http://www.cs.technion.ac.il/~gabr/papers/accio.pdf}, abstract = {Although text categorization is a burgeoning area of IR research, readily available test collections in this field are surprisingly scarce. We describe a methodology and system (named ACCIO) for automatically acquiring labeled datasets for text categorization from the World Wide Web, by capitalizing on the body of knowledge encoded in the structure of existing hierarchical directories such as the Open Directory. We define parameters of categories that make it possible to acquire numerous datasets with desired properties, which in turn allow better control over categorization experiments. In particular, we develop metrics that estimate the difficulty of a dataset by examining the host directory structure. These metrics are shown to be good predictors of categorization accuracy that can be achieved on a dataset, and serve as efficient heuristics for generating datasets subject to user's requirements. A large collection of automatically generated datasets are made available for other researchers to use.}, } @inProceedings{Debole03, author = {Franca Debole and Fabrizio Sebastiani}, title = {Supervised term weighting for automated text categorization}, year = {2003}, booktitle = {Proceedings of SAC-03, 18th ACM Symposium on Applied Computing}, address = {Melbourne, US}, publisher = {ACM Press, New York, US}, pages = {784--788}, url = {http://www.math.unipd.it/~fabseb60/Publications/SAC03b.pdf}, note = {An extended version appears as~\cite{Debole04a}}, abstract = {The construction of a text classifier usually involves (i) a phase of \emph{term selection}, in which the most relevant terms for the classification task are identified, (ii) a phase of \emph{term weighting}, in which document weights for the selected terms are computed, and (iii) a phase of \emph{classifier learning}, in which a classifier is generated from the weighted representations of the training documents. This process involves an activity of {\em supervised learning}, in which information on the membership of training documents in categories is used. Traditionally, supervised learning enters only phases (i) and (iii). In this paper we propose instead that learning from training data should also affect phase (ii), i.e.\ that information on the membership of training documents to categories be used to determine term weights. We call this idea \emph{supervised term weighting} (STW). As an example, we propose a number of ``supervised variants'' of $tfidf$ weighting, obtained by replacing the $idf$ function with the function that has been used in phase (i) for term selection. We present experimental results obtained on the standard \textsf{Reuters-21578} benchmark with one classifier learning method (support vector machines), three term selection functions (information gain, chi-square, and gain ratio), and both local and global term selection and weighting.}, } @inProceedings{Debole04, author = {Franca Debole and Fabrizio Sebastiani}, title = {An Analysis of the Relative Difficulty of Reuters-21578 Subsets}, year = {2004}, booktitle = {Proceedings of LREC-04, 4th International Conference on Language Resources and Evaluation}, address = {Lisbon, PT}, pages = {}, url = {http://www.math.unipd.it/~fabseb60/Publications/LREC04.pdf}, abstract = {The existence, public availability, and widespread acceptance of a standard benchmark for a given information retrieval (IR) task are beneficial to research on this task, since they allow different researchers to experimentally compare their own systems by comparing the results they have obtained on this benchmark. The \textsf{Reuters-21578} test collection, together with its earlier variants, has been such a standard benchmark for the text categorization (TC) task throughout the last ten years. However, the benefits that this has brought about have somehow been limited by the fact that different researchers have ``carved'' different subsets out of this collection, and tested their systems on one of these subsets only; systems that have been tested on different \textsf{Reuters-21578} subsets are thus not readily comparable. In this paper we present a systematic, comparative experimental study of the three subsets of \textsf{Reuters-21578} that have been most popular among TC researchers. The results we obtain allow us to determine the relative difficulty of these subsets, thus establishing an indirect means for comparing TC systems that have, or will be, tested on these different subsets.}, note = {Forthcoming}, } @inCollection{Debole04a, author = {Franca Debole and Fabrizio Sebastiani}, title = {Supervised Term Weighting for Automated Text Categorization}, year = {2004}, booktitle = {Text Mining and its Applications}, editor = {Spiros Sirmakessis}, publisher = {Physica-Verlag, Heidelberg, DE}, series = {Number 138 in the ``Studies in Fuzziness and Soft Computing'' series}, pages = {81--98}, url = {http://www.math.unipd.it/~fabseb60/Publications/NEMIS04.pdf}, abstract = {The construction of a text classifier usually involves (i) a phase of \emph{term selection}, in which the most relevant terms for the classification task are identified, (ii) a phase of \emph{term weighting}, in which document weights for the selected terms are computed, and (iii) a phase of \emph{classifier learning}, in which a classifier is generated from the weighted representations of the training documents. This process involves an activity of {\em supervised learning}, in which information on the membership of training documents in categories is used. Traditionally, supervised learning enters only phases (i) and (iii). In this paper we propose instead that learning from the training data should also affect phase (ii), i.e.\ that information on the membership of training documents to categories be used to determine term weights. We call this idea \emph{supervised term weighting} (STW). As an example of STW, we propose a number of ``supervised variants'' of $tfidf$ weighting, obtained by replacing the $idf$ function with the function that has been used in phase (i) for term selection. The use of STW allows the terms that are distributed most differently in the positive and negative examples of the categories of interest to be weighted highest. We present experimental results obtained on the standard \textsf{Reuters-21578} benchmark with three classifier learning methods (Rocchio, $k$-NN, and support vector machines), three term selection functions (information gain, chi-square, and gain ratio), and both local and global term selection and weighting.}, } @inProceedings{Debole04c, author = {Franca Debole and Fabrizio Sebastiani}, title = {An Analysis of the Relative Hardness of Reuters-21578 Subsets}, year = {2004}, booktitle = {Proceedings of LREC-04, 4th International Conference on Language Resources and Evaluation}, address = {Lisbon, PT}, pages = {971--974}, url = {http://www.math.unipd.it/~fabseb60/Publications/LREC04.pdf}, } @article{Debole05, author = {Franca Debole and Fabrizio Sebastiani}, title = {An Analysis of the Relative Hardness of Reuters-21578 Subsets}, journal = {Journal of the American Society for Information Science and Technology}, year = {2004}, volume = {56}, number = {6}, pages = {584--596}, url = {http://www.math.unipd.it/~fabseb60/Publications/JASIST05.pdf}, abstract = {The existence, public availability, and widespread acceptance of a standard benchmark for a given information retrieval (IR) task are beneficial to research on this task, since they allow different researchers to experimentally compare their own systems by comparing the results they have obtained on this benchmark. The \textsf{Reuters-21578} test collection, together with its earlier variants, has been such a standard benchmark for the text categorization (TC) task throughout the last ten years. However, the benefits that this has brought about have somehow been limited by the fact that different researchers have ``carved'' different subsets out of this collection, and tested their systems on one of these subsets only; systems that have been tested on different \textsf{Reuters-21578} subsets are thus not readily comparable. In this paper we present a systematic, comparative experimental study of the three subsets of \textsf{Reuters-21578} that have been most popular among TC researchers. The results we obtain allow us to determine the relative hardness of these subsets, thus establishing an indirect means for comparing TC systems that have, or will be, tested on these different subsets.}, } @inProceedings{deBuenaga97, author = {De Buenaga Rodr{\'{\i}}guez, Manuel and G{\'o}mez-Hidalgo, Jos{\'e} Mar{\'{\i}}a and D{\'{\i}}az-Agudo, Bel{\'e}n}, title = {Using {WordNet} to Complement Training Information in Text Categorization}, booktitle = {Proceedings of RANLP-97, 2nd International Conference on Recent Advances in Natural Language Processing}, publisher = {}, editor = {Ruslan Milkov and Nicolas Nicolov and Nilokai Nikolov}, address = {Tzigov Chark, BL}, pages = {}, year = {1997}, url = {http://xxx.unizar.es/ps/cmp-lg/9709007}, abstract = {Automatic Text Categorization (TC) is a complex and useful task for many natural language applications, and is usually performed through the use of a set of manually classified documents, a training collection. We suggest the utilization of additional resources like lexical databases to increase the amount of information that TC systems make use of, and thus, to improve their performance. Our approach integrates WordNet information with two training approaches through the Vector Space Model. The training approaches we test are the Rocchio (relevance feedback) and the Widrow-Hoff (machine learning) algorithms. Results obtained from evaluation show that the integration of WordNet clearly outperforms training approaches, and that an integrated technique can effectively address the classification of low frequency categories.}, } @inProceedings{deLima98, author = {De Lima, Luciano R. and Laender, Alberto H. and Ribeiro-Neto, Berthier A.}, title = {A hierarchical approach to the automatic categorization of medical documents}, booktitle = {Proceedings of CIKM-98, 7th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {Georges Gardarin and James C. French and Niki Pissinou and Kia Makki and Luc Bouganim}, year = {1998}, address = {Bethesda, US}, pages = {132--139}, url = {http://www.acm.org/pubs/articles/proceedings/cikm/288627/p132-de_lima/p132-de_lima.pdf}, abstract = {}, } @inProceedings{Denoyer01, author = {Ludovic Denoyer and Hugo Zaragoza and Patrick Gallinari}, title = {HMM-based Passage Models for Document Classification and Ranking}, booktitle = {Proceedings of ECIR-01, 23rd European Colloquium on Information Retrieval Research}, editor = {}, year = {2001}, address = {Darmstadt, DE}, publisher = {}, pages = {126--135}, url = {http://www-connex.lip6.fr/~denoyer/publications/denoyer-final-ecir01.ps}, abstract = {We present an application of Hidden Markov Models to supervised document classification and ranking. We consider a family of models that take into account the fact that relevant documents may contain irrelevant passages; the originality of the model is that it does not explicitly segment documents but rather considers all possible segmentations in its final score. This model generalizes the multinomial Naive Bayes and it is derived from a more general model for different access tasks. The model is evaluated on the REUTERS test collection and compared to the multinomial Naive Bayes model. It is shown to be more robust with respect to the training set size and to improve the performance both for ranking and classification, specially for classes with few training examples.}, } @inProceedings{Denoyer03, author = {Ludovic Denoyer and Patrick Gallinari}, title = {A Belief Networks-Based Generative Model for Structured Documents. An Application to the XML Categorization}, booktitle = {Proceedings of MLDM-03, 3rd International Conference on Machine Learning and Data Mining in Pattern Recognition}, editor = {Petra Perner and Azriel Rosenfeld}, year = {2003}, address = {Leipzig, DE}, publisher = {Springer Verlag, Heidelberg, DE}, pages = {328--342}, url = {http://www.springerlink.com/openurl.asp?genre=article&issn=0302-9743&volume=2734&spage=328}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2734}, abstract = {We present a generative Bayesian model for the modeling of structured (e.g. XML) documents. This model allows us to simultaneously take into account structure and content information. It is used here for classifying XML documents. We adopt a machine learning approach and the model parameters are learned from a labeled training set of representative documents. We discuss the role of structural information for classification and describe experiments on a small collection of class labeled structured documents. We also present preliminary results showing how this model could classify documents with DTDs not represented in the training set.}, } @inProceedings{Denoyer03a, author = {Ludovic Denoyer and Jean-No{\"{e}}l Vittaut and Patrick Gallinari and Sylvie Brunessaux and Stephan Brunessaux}, title = {Structured multimedia document classification}, booktitle = {Proceedings of DOCENG-03, ACM Symposium on Document engineering}, publisher = {ACM Press, New York, US}, editor = {}, year = {2003}, address = {Grenoble, FR}, pages = {153--160}, url = {http://doi.acm.org/10.1145/958220.958249}, abstract = {We propose a new statistical model for the classification of structured documents and consider its use for multimedia document classification. Its main originality is its ability to simultaneously take into account the structural and the content information present in a structured document, and also to cope with different types of content (text, image, etc). We present experiments on the classification of multilingual pornographic HTML pages using text and image data. The system accurately classifies porn sites from 8 European languages. This corpus has been developed by EADS company in the context of a large Web site filtering application.}, } @article{Denoyer04, author = {Ludovic Denoyer and Patrick Gallinari}, title = {Bayesian network model for semi-structured document classification}, journal = {Information Processing and Management}, year = {2004}, volume = {40}, number = {5}, pages = {807--827}, url = {}, abstract = {}, } @article{deVel01, author = {Olivier Y. {De Vel} and Alison Anderson and Malcolm Corney and George M. Mohay}, title = {Mining Email Content for Author Identification Forensics}, journal = {SIGMOD Record}, year = {2001}, volume = {30}, number = {4}, pages = {55--64}, url = {}, abstract = {We describe an investigation into e-mail content mining for author identification, or authorship attribution, for the purpose of forensic investigation. We focus our discussion on the ability to discriminate between authors for the case of both aggregated e-mail topics as well as across different email topics. An extended set of e-mail document features including structural characteristics and linguistic patterns were derived and, together with a Support Vector Machine learning algorithm, were used for mining the e-mail content. Experiments using a number of e-mail documents generated by different authors on a set of topics gave promising results for both aggregated and multi-topic author categorisation.}, } @inProceedings{Dhillon02, author = {Inderjit Dhillon and Subramanyam Mallela and Rahul Kumar}, title = {Enhanced word clustering for hierarchical text classification}, booktitle = {Proceedings of KDD-02, 8th ACM International Conference on Knowledge Discovery and Data Mining}, publisher = {ACM Press, New York, US}, editor = {}, year = {2002}, address = {Edmonton, CA}, pages = {191--200}, url = {}, abstract = {In this paper we propose a new information-theoretic divisive algorithm for word clustering applied to text classification. In previous work, such "distributional clustering" of features has been found to achieve improvements over feature selection in terms of classification accuracy, especially at lower number of features [2, 28]. However the existing clustering, techniques are agglomerative in nature and result in (i) suboptimal word clusters and (ii) high computational cost. In order to explicitly capture the optimality of word clusters in an information theoretic framework, we first derive a global criterion for feature clustering. We then present a fast, divisive algorithm that monotonically decreases this objective function value, thus converging to a local minimum. We show that our algorithm minimizes the "within-cluster Jensen-Shannon divergence" while simultaneously maximizing the "between-cluster Jensen-Shannon divergence". In comparison to the previously proposed agglomerative strategies our divisive algorithm achieves higher classification accuracy especially at lower number of features. We further show that feature clustering is an effective technique for building smaller class models in hierarchical classification. We present detailed experimental results using Naive Bayes and Support Vector Machines on the 20 Newsgroups data set and a 3-level hierarchy of HTML documents collected from Dmoz Open Directory.}, } @article{Dhillon03, author = {Inderjit Dhillon and Subramanyam Mallela and Rahul Kumar}, title = {A divisive information-theoretic feature clustering algorithm for text classification}, journal = {Journal of Machine Learning Research}, volume = {3}, month = {March}, pages = {1265--1287}, year = {2003}, url = {http://www.jmlr.org/papers/volume3/dhillon03a/dhillon03a.pdf}, abstract = {High dimensionality of text can be a deterrent in applying complex learners such as Support Vector Machines to the task of text classification. Feature clustering is a powerful alternative to feature selection for reducing the dimensionality of text data. In this paper we propose a new information-theoretic divisive algorithm for feature/word clustering and apply it to text classification. Existing techniques for such "distributional clustering" of words are agglomerative in nature and result in (i) sub-optimal word clusters and (ii) high computational cost. In order to explicitly capture the optimality of word clusters in an information theoretic framework, we first derive a global criterion for feature clustering. We then present a fast, divisive algorithm that monotonically decreases this objective function value. We show that our algorithm minimizes the "within-cluster Jensen-Shannon divergence" while simultaneously maximizing the "between-cluster Jensen-Shannon divergence". In comparison to the previously proposed agglomerative strategies our divisive algorithm is much faster and achieves comparable or higher classification accuracies. We further show that feature clustering is an effective technique for building smaller class models in hierarchical classification. We present detailed experimental results using Naive Bayes and Support Vector Machines on the 20Newsgroups data set and a 3-level hierarchy of HTML documents collected from the Open Directory project (www.dmoz.org).}, } @inProceedings{Diao00, author = {Yanlei Diao and Hongjun Lu and Dekai Wu}, title = {A comparative study of classification-based personal e-mail filtering}, booktitle = {Proceedings of PAKDD-00, 4th Pacific-Asia Conference on Knowledge Discovery and Data Mining}, editor = {Takao Terano and Huan Liu and Arbee L.P. Chen}, pages = {408--419}, year = {2000}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Kyoto, JP}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1805}, url = {http://www.cs.berkeley.edu/~diaoyl/publications/pakdd00.ps}, abstract = {This paper addresses personal E-mail filtering by casting it in the framework of text classification. Modeled as semi-structured documents, E-mail messages consist of a set of fields with predefined semantics and a number of variable length free-text fields. While most work on classification either concentrates on structured data or free text, the work in this paper deals with both of them. To perform classification, a naive Bayesian classifier was designed and implemented, and a decision tree based classifier was implemented. The design considerations and implementation issues are discussed. Using a relatively large amount of real personal E-mail data, a comprehensive comparative study was conducted using the two classifiers. The importance of different features is reported. Results of other issues related to building an effective personal E-mail classifier are presented and discussed. It is shown that both classifiers can perform filtering with reasonable accuracy. While the decision tree based classifier outperforms the Bayesian classifier when features and training size are selected optimally for both, a carefully designed naive Bayesian classifier is more robust.}, } @inProceedings{Diaz98, author = {D{\'{\i}}az Esteban, Alberto and De Buenaga Rodr{\'{\i}}guez, Manuel and Ure{\~n}a L{\'o}pez, L. Alfonso and Garc{\'{\i}}a Vega, Manuel}, title = {Integrating Linguistic Resources in an Uniform Way for Text Classification Tasks}, booktitle = {Proceedings of LREC-98, 1st International Conference on Language Resources and Evaluation}, publisher = {}, editor = {Antonio Rubio and Natividad Gallardo and Rosa Castro and Antonio Tejada}, address = {Grenada, ES}, pages = {1197--1204}, year = {1998}, url = {http://www.esi.uem.es/laboratorios/sinai/postscripts/lrec98.ps}, abstract = {Applications based on automatic text classification tasks, like text categorization (TC), word sense disambiguation (WSD), text filtering or routing, monolingual or multilingual information retrieval, and text summarization could obtain serious improvements by integrating linguistic resources in the current methods. We present an approach using the Vector Space Model to integrate two different kind of resources: a lexical database and training collections, in text content analysis tasks. The training approaches we test are the Rocchio (relevance feedback) and the Widrow-Hoff (machine learning) algorithms and WordNet as the lexical database. We have delevoped experimental systems for TC and WSD. Results obtained from evaluation show that the integration of WordNet can outperform approaches based only on training.}, } @article{Diederich03, author = {Diederich, Joachim and Kindermann, J{\"{o}}rg and Leopold, Edda and Paass, Gerhard}, title = {Authorship Attribution with Support Vector Machines}, journal = {Applied Intelligence}, year = {2003}, volume = {19}, number = {1/2}, pages = {109--123}, url = {http://ipsapp007.kluweronline.com/content/getfile/4504/36/6/abstract.htm}, abstract = {In this paper we explore the use of text-mining methods for the identification of the author of a text. We apply the support vector machine (SVM) to this problem, as it is able to cope with half a million of inputs it requires no feature selection and can process the frequency vector of all words of a text. We performed a number of experiments with texts from a German newspaper. With nearly perfect reliability the SVM was able to reject other authors and detected the target author in 60-80\% of the cases. In a second experiment, we ignored nouns, verbs and adjectives and replaced them by grammatical tags and bigrams. This resulted in slightly reduced performance. Author detection with SVMs on full word forms was remarkably robust even if the author wrote about different topics.}, } @inProceedings{DiNunzio03, author = {Giorgio M. {Di Nunzio} and Alessandro Micarelli}, title = {Does a New Simple Gaussian Weighting Approach Perform Well in Text Categorization?}, booktitle = {Proceedings of IJCAI-03, 18th International Joint Conference on Artificial Intelligence}, editor = {Georg Gottlob and Toby Walsh}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, year = {2003}, address = {Acapulco, MX}, pages = {581--586}, url = {}, abstract = {A new approach to the Text Categorization problem is here presented. It is called Gaussian Weighting and it is a supervised learning algorithm that, during the training phase, estimates two very simple and easily computable statistics which are: the Presence \emphP, how much a term \emph{t} is present in a category \emph{c}; the Expressiveness \emphE, how much \emph{t} is present outside \emph{c} in the rest of the domain. Once the system has learned this information, a Gaussian function is shaped for each term of a category, in order to assign the term a weight that estimates the level of its importance for that particular category. We tested our learning method on the task of single-label classification using the Reuters-21578 benchmark. The outcome of the result was quite impressive: in different experimental setups, we reached a micro-averaged F1-measure of 0.89, with a peak of 0.899. Moreover, a macro-averaged Recall and Precision was calculated: the former reported a 0.72, the latter a 0.79. These results reach most of the state-of-the-art techniques of machine learning applied to Text Categorization, demonstrating that this new weighting scheme does perform well on this particular task.}, } @inProceedings{DiNunzio04, author = {Giorgio M. {Di Nunzio}}, title = {A Bidimensional View of Documents for Text Categorisation}, booktitle = {Proceedings of ECIR-04, 26th European Conference on Information Retrieval Research}, editor = {Sharon McDonald and John Tait}, year = {2004}, address = {Sunderland, UK}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2997}, pages = {112--126}, url = {http://springerlink.metapress.com/openurl.asp?genre=article&issn=0302-9743&volume=2997&spage=112}, abstract = {The question addressed in this paper is to find a bidimensional representation of textual documents for the problem of text categorisation. The projection of documents is performed following subsequent steps. The main idea is to consider a possible double aspect of the importance of a word: the local importance in a category, and the global importance in the rest of the categories. This information is combined properly and summarized in two coordinates. Then, a machine learning method may be used in this simple bidimensional space to classify the documents. The results that can be obtained in this space are satisfactory with respect to the best state-of-the-art performances.}, } @inProceedings{Dorre99, author = {Jochen D{\"o}rre and Peter Gerstl and Roland Seiffert}, title = {Text mining: finding nuggets in mountains of textual data}, booktitle = {Proceedings of KDD-99, 5th ACM International Conference on Knowledge Discovery and Data Mining}, publisher = {ACM Press, New York, US}, editor = {}, year = {1999}, address = {San Diego, US}, pages = {398--401}, url = {http://www.acm.org/pubs/articles/proceedings/ai/312129/p398-dorre/p398-dorre.pdf}, abstract = {Text mining applies the same analytical functions of data mining to the domain of textual information, relying on sophisticated text analysis techniques that distill information from free-text documents. IBM's Intelligent Miner for Text provides the necessary tools to unlock the business information that is ''trapped'' in email, insurance claims, news feeds, or other document repositories. It has been successfully applied in analyzing patent portfolios, customer complaint letters, and even competitors' Web pages. After defining our notion of ``text mining'', we focus on the differences between text and data mining and describe in some more detail the unique technologies that are key to successful text mining.}, } @article{Doyle65, author = {Lauren B. Doyle}, title = {Is automatic classification a reasonable application of statistical analysis of text?}, journal = {Journal of the ACM}, volume = {12}, number = {4}, year = {1965}, pages = {473--489}, url = {http://www.acm.org/pubs/articles/journals/jacm/1965-12-4/p473-doyle/p473-doyle.pdf}, abstract = {The statistical approach to the analysis of document collections and retrieval therefrom has proceeded along two main lines, associative machine searching and automatic classification. The former approach has been favored because of the tendency of people in the computer field to strive for new methods of dealing with the literature -- methods which do not resemble those of traditional libraries. But automatic classification study also has been thriving; some of the reasons for this are discussed. The crucial question of the quality of automatic classification is treated at considerable length, and empirical data are introduced to support the hypothesis that classification quality improves as more information about each document is used for input to the classification program. Six nonjudgmental criteria are used in testing the hypothesis for 100 keyword lists (each list representing a document) for a series of computer runs in which the number of words per document is increased progressively from 12 to 36. Four of the six criteria indicate the hypothesis holds, and two point to no effect. Previous work of this kind has been confined to the range of one through eight words per document. Finally, the future of automatic classification and some of the practical problems to be faced are outlined.}, } @article{Drucker99, author = {Harris Drucker and Vladimir Vapnik and Dongui Wu}, title = {Support vector machines for spam categorization}, journal = {IEEE Transactions on Neural Networks}, year = {1999}, number = {5}, volume = {10}, pages = {1048--1054}, url = {http://www.monmouth.edu/~drucker/SVM_spam_article_compete.PDF}, abstract = {We study the use of Support Vector Machines (SVMs) in classifying email as spam or nonspam by comparing it to three other classification algorithms: Ripper, Rocchio, and boosting decision trees. These four algorithms were tested on two different data sets: one data set where the number of features were constrained to the 1000 best features and another data set where the dimensionality was over 7000. SVMs performed best when using binary features. For both data sets, boosting trees and SVMs had acceptable test performance in terms of accuracy and speed. However, SVMs had significantly less training time.} } @inProceedings{Dumais:2000:HCW, author = {Susan T. Dumais and Hao Chen}, title = {Hierarchical classification of Web content}, booktitle = {Proceedings of SIGIR-00, 23rd ACM International Conference on Research and Development in Information Retrieval}, editor = {Nicholas J. Belkin and Peter Ingwersen and Mun-Kew Leong}, publisher = {ACM Press, New York, US}, address = {Athens, GR}, year = {2000}, pages = {256--263}, url = {http://research.microsoft.com/~sdumais/sigir00.pdf}, abstract = {This paper explores the use of hierarchical structure for classifying a large, heterogeneous collection of web content. The hierarchical structure is initially used to train different second-level classifiers. In the hierarchical case, a model is learned to distinguish a second-level category from other categories within the same top level. In the flat non-hierarchical case, a model distinguishes a second-level category from all other second-level categories. Scoring rules can further take advantage of the hierarchy by considering only second-level categories that exceed a threshold at the top level. We use support vector machine (SVM) classifiers, which have been shown to be efficient and effective for classification, but not previously explored in the context of hierarchical classification. We found small advantages in accuracy for hierarchical models over flat models. For the hierarchical approach, we found the same accuracy using a sequential Boolean decision rule and a multiplicative decision rule. Since the sequential approach is much more efficient, requiring only 14\%-16\% of the comparisons used in the other approaches, we find it to be a good choice for classifying text into large hierarchical structures.} } @inProceedings{Dumais98, author = {Susan T. Dumais and John Platt and David Heckerman and Mehran Sahami}, title = {Inductive learning algorithms and representations for text categorization}, booktitle = {Proceedings of CIKM-98, 7th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {Georges Gardarin and James C. French and Niki Pissinou and Kia Makki and Luc Bouganim}, year = {1998}, address = {Bethesda, US}, pages = {148--155}, url = {http://robotics.stanford.edu/users/sahami/papers-dir/cikm98.pdf}, abstract = {Text categorization - the assignment of natural language texts to one or more predefined categories based on their content - is an important component in many information organization and management tasks. We compare the effectiveness of five different automatic learning algorithms for text categorization in terms of learning speed, real-time classification speed, and classification accuracy. We also examine training set size, and alternative document representations. Very accurate text classifiers can be learned automatically from training examples. Linear Support Vector Machines (SVMs) are particularly promising because they are very accurate, quick to train, and quick to evaluate.} } @inProceedings{ElYaniv01, author = {Ran El-Yaniv and Oren Souroujon}, title = {Iterative Double Clustering for Unsupervised and Semi-supervised Learning}, booktitle = {Proceedings of ECML-01, 12th European Conference on Machine Learning}, editor = {Luc De Raedt and Peter A. Flach}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Freiburg, DE}, year = {2001}, pages = {121--132}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2167}, url = {http://link.springer.de/link/service/series/0558/papers/2167/21670121.pdf}, abstract = {This paper studies the Iterative Double Clustering (IDC) meta-clustering algorithm, a new extension of the recent Double Clustering (DC) method of Slonim and Tishby that exhibited impressive performance on text categorization tasks. Using synthetically generated data we empirically demonstrate that whenever the DC procedure is successful in recovering some of the structure hidden in the data, the extended IDC procedure can incrementally compute a dramatically better classification, with minor additional computational resources. We demonstrate that the IDC algorithm is especially advantageous when the data exhibits high attribute noise. Our simulation results also show the effectiveness of IDC in text categorization problems. Surprisingly, this unsupervised procedure can be competitive with a (supervised) SVM trained with a small training set. Finally, we propose a natural extension of IDC for (semi-supervised) transductive learning where we are given both labeled and unlabeled examples, and present preliminary empirical results showing the plausibility of the extended method in a semi-supervised setting.}, } @inProceedings{Escudero00, author = {Gerard Escudero and Llu{\'{\i}}s M{\`{a}}rquez and German Rigau}, title = {Boosting applied to word sense disambiguation}, booktitle = {Proceedings of ECML-00, 11th European Conference on Machine Learning}, editor = {Ramon L{\'{o}}pez De M{\'{a}}ntaras and Enric Plaza}, address = {Barcelona, ES}, pages = {129--141}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1810}, year = {2000}, url = {http://www.lsi.upc.es/~escudero/recerca/ecml00.pdf}, abstract = {In this paper Schapire and Singer's AdaBoost.MH boosting algorithm is applied to the Word Sense Disambiguation (WSD) problem. Initial experiments on a set of 15 selected polysemous words show that the boosting approach surpasses Naive Bayes and Exemplar-based approaches, which represent state-of-the-art accuracy on supervised WSD. In order to make boosting practical for a real learning domain of thousands of words, several ways of accelerating the algorithm by reducing the feature space are studied. The best variant, which we call LazyBoosting, is tested on the largest sense-tagged corpus available containing 192,800 examples of the 191 most frequent and ambiguous English words. Again, boosting compares favourably to the other benchmark algorithms.}, } @article{Fall03, author = {C. J. Fall and A. T{\"o}rcsv{\'a}ri and K. Benzineb and G. Karetka}, title = {Automated Categorization in the {International Patent Classification}}, journal = {SIGIR Forum}, year = {2003}, pages = {10--25}, volume = {37}, number = {1}, url = {http://www.acm.org/sigir/forum/S2003/CJF_Manuscript_sigir.pdf}, abstract = {A new reference collection of patent documents for training and testing automated categorization systems is established and described in detail. This collection is tailored for automating the attribution of international patent classification codes to patent applications and is made publicly available for future research work. We report the results of applying a variety of machine learning algorithms to the automated categorization of English-language patent documents. This procedure involves a complex hierarchical taxonomy, within which we classify documents into 114 classes and 451 subclasses. Several measures of categorization success are described and evaluated. We investigate how best to resolve the training problems related to the attribution of multiple classification codes to each patent document.}, } @inProceedings{Fangmeyer68, author = {Hermann Fangmeyer and Gerhard Lustig}, title = {The EURATOM automatic indexing project}, booktitle = {Proceedings of the IFIP Congress (Booklet J)}, publisher = {}, editor = {}, year = {1968}, address = {Edinburgh, UK}, pages = {66--70}, url = {}, abstract = {}, } @inProceedings{Fangmeyer70, author = {Hermann Fangmeyer and Gerhard Lustig}, title = {Experiments with the CETIS automated indexing system}, booktitle = {Proceedings of the Symposium on the Handling of Nuclear Information}, publisher = {International Atomic Energy Agency}, editor = {}, year = {1970}, address = {}, pages = {557--567}, url = {}, abstract = {}, } @inProceedings{Ferilli01, author = {Stefano Ferilli and Nicola Fanizzi and Gianni Semeraro}, title = {Learning logic models for automated text categorization}, booktitle = {Proceedings of AI*IA-01, 7th Congress of the Italian Association for Artificial Intelligence}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2175}, editor = {Floriana Esposito}, year = {2001}, pages = {81--86}, address = {Bari, IT}, url = {http://link.springer.de/link/service/series/0558/papers/2175/21750081.pdf}, abstract = {This work addresses a logical approach to text categorization inside a framework aimed at full automatic paper document processing. The logic representation of sentences required by the adopted learning algorithm is obtained by detecting structure in raw text trough a parser. A preliminary experimentation proved that the logic approach is able to capture the semantics underlying some kind of sentences, even if the assessment of the efficiency of such a method, as well as a comparison with other related approaches, has still to be carried out.}, } @article{Field75, author = {B.J. Field}, title = {Towards automatic indexing: automatic assignment of controlled-language indexing and classification from free indexing}, year = {1975}, journal = {Journal of Documentation}, volume = {31}, number = {4}, pages = {246--265}, url = {}, abstract = {}, } @inProceedings{Finn02, author = {Aidan Finn and Nicholas Kushmerick and Barry Smyth}, title = {Genre Classification and Domain Transfer for Information Filtering}, booktitle = {Proceedings of ECIR-02, 24th European Colloquium on Information Retrieval Research}, editor = {Fabio Crestani and Mark Girolami and Van Rijsbergen, Cornelis J.}, year = {2002}, address = {Glasgow, UK}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2291}, pages = {353--362}, url = {http://www.cs.ucd.ie/staff/nick/home/research/download/finn-ecir2002.ps.gz}, abstract = {The World Wide Web is a vast repository of information, but the sheer volume makes it difficult to identify useful documents. We identify document genre is an important factor in retrieving useful documents and focus on the novel document genre dimension of subjectivity. We investigate three approaches to automatically classifying documents by genre: traditional bag of words techniques, part-of-speech statistics, and hand-crafted shallow linguistic features. We are particularly interested in domain transfer: how well the learned classifiers generalize from the training corpus to a new document corpus. Our experiments demonstrate that the part-of-speech approach is better than traditional bag of words techniques, particularly in the domain transfer conditions.}, } @inProceedings{Fisher03, author = {Michelle Fisher and Richard Everson}, title = {When are links useful? Experiments in text classification}, booktitle = {Proceedings of ECIR-03, 25th European Conference on Information Retrieval}, publisher = {Springer Verlag}, editor = {Fabrizio Sebastiani}, address = {Pisa, IT}, year = {2003}, pages = {41--56}, url = {http://link.springer.de/link/service/series/0558/papers/2633/26330041.pdf}, abstract = {Link analysis methods have become popular for information access tasks, especially information retrieval, where the link information in a document collection is used to complement the traditionally used content information. However, there has been little firm evidence to confirm the utility of link information. We show that link information can be useful when the document collection has a sufficiently high link density and links are of sufficiently high quality. We report experiments on text classification of the Cora and WebKB data sets using Probabilistic Latent Semantic Analysis and Probabilistic Hypertext Induced Topic Selection. Comparison with manually assigned classes shows that link information enhances classification in data with sufficiently high link density, but is detrimental to performance at low link densities or if the quality of the links is degraded. We introduce a new frequency-based method for selecting the most useful citations from a document collection for use in the model.}, } @inProceedings{Forman02, author = {George Forman}, title = {Choose Your Words Carefully: An Empirical Study of Feature Selection Metrics for Text Classification}, booktitle = {Proceedings of PKDD-02, 6th European Conference on Principles of Data Mining and Knowledge Discovery}, editor = {Tapio Elomaa and Heikki Mannila and H. Toivonen}, address = {Helsinki, FI}, pages = {150--162}, year = {2002}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2431}, url = {http://springerlink.metapress.com/openurl.asp?genre=article&issn=0302-9743&volume=2431&spage=150}, abstract = {Good feature selection is essential for text classification to make it tractable for machine learning, and to improve classification performance. This study benchmarks the performance of twelve feature selection metrics across 229 text classification problems drawn from Reuters, OHSUMED, TREC, etc. using Support Vector Machines. The results are analyzed for various objectives. For best accuracy, F-measure or recall, the findings reveal an outstanding new feature selection metric, "Bi-Normal Separation" (BNS). For precision alone, however, Information Gain (IG) was superior. A new evaluation methodology is offered that focuses on the needs of the data mining practitioner who seeks to choose one or two metrics to try that are mostly likely to have the best performance for the single dataset at hand. This analysis determined, for example, that IG and Chi-Squared have correlated failures for precision, and that IG paired with BNS is a better choice.}, } @article{Forman03, author = {George Forman}, title = {An Extensive Empirical Study of Feature Selection Metrics for Text Classification}, journal = {Journal of Machine Learning Research}, volume = {3}, month = {March}, pages = {1289--1305}, year = {2003}, url = {http://www.jmlr.org/papers/v3/forman03a.html}, abstract = {Machine learning for text classification is the cornerstone of document categorization, news filtering, document routing, and personalization. In text domains, effective feature selection is essential to make the learning task efficient and more accurate. This paper presents an empirical comparison of twelve feature selection methods (e.g.\ Information Gain) evaluated on a benchmark of 229 text classification problem instances that were gathered from Reuters, TREC, OHSUMED, etc. The results are analyzed from multiple goal perspectives-accuracy, F-measure, precision, and recall-since each is appropriate in different situations. The results reveal that a new feature selection metric we call 'Bi-Normal Separation' (BNS), outperformed the others by a substantial margin in most situations. This margin widened in tasks with high class skew, which is rampant in text classification problems and is particularly challenging for induction algorithms. A new evaluation methodology is offered that focuses on the needs of the data mining practitioner faced with a single dataset who seeks to choose one (or a pair of) metrics that are most likely to yield the best performance. From this perspective, BNS was the top single choice for all goals except precision, for which Information Gain yielded the best result most often. This analysis also revealed, for example, that Information Gain and Chi-Squared have correlated failures, and so they work poorly together. When choosing optimal pairs of metrics for each of the four performance goals, BNS is consistently a member of the pair---e.g., for greatest recall, the pair BNS + F1-measure yielded the best performance on the greatest number of tasks by a considerable margin.}, } @inProceedings{Forman04, author = {George Forman}, title = {A pitfall and solution in multi-class feature selection for text classification}, booktitle = {Proceedings of ICML-04, 21st International Conference on Machine Learning}, editor = {Carla E. Brodley}, year = {2004}, address = {Banff, CA}, pages = {}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://doi.acm.org/10.1145/1015330.1015356}, abstract = {Information Gain is a well-known and empirically proven method for high-dimensional feature selection. We found that it and other existing methods failed to produce good results on an industrial text classification problem. On investigating the root cause, we find that a large class of feature scoring methods suffers a pitfall: they can be blinded by a surplus of strongly predictive features for some classes, while largely ignoring features needed to discriminate difficult classes. In this paper we demonstrate this pitfall hurts performance even for a relatively uniform text classification task. Based on this understanding, we present solutions inspired by round-robin scheduling that avoid this pitfall, without resorting to costly wrapper methods. Empirical evaluation on 19 datasets shows substantial improvements.}, } @inProceedings{Forman04a, author = {George Forman and Ira Cohen}, title = {Learning from little: comparison of classifiers given little training}, booktitle = {Proceedings of PKDD-04, 8th European Conference on Principles of Data Mining and Knowledge Discovery}, editor = {Jean-Fran{\c{c}}ois Boulicaut and Floriana Esposito and Fosca Giannotti and Dino Pedreschi}, address = {Pisa, IT}, pages = {161--172}, year = {2004}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 3202}, url = {}, abstract = {}, } @inCollection{Forsyth99, author = {Richard S. Forsyth}, title = {New directions in text categorization}, editor = {Alex Gammerman}, booktitle = {Causal models and intelligent data management}, publisher = {Springer Verlag}, address = {Heidelberg, DE}, year = {1999}, pages = {151--185}, url = {}, abstract = {}, } @inProceedings{Frank00, author = {Eibe Frank and Chang Chui and Ian H. Witten}, title = {Text Categorization Using Compression Models}, booktitle = {Proceedings of DCC-00, IEEE Data Compression Conference}, editor = {Storer, James A. and Cohn, Martin}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, year = {2000}, address = {Snowbird, US}, pages = {200--209}, url = {http://dlib.computer.org/conferen/dcc/0592/pdf/05920555.pdf}, abstract = {Text categorization is the assignment of natural language texts to predefined categories based on their content. It has often been observed that compression seems to provide a very promising approach to categorization. The overall compression of an article with respect to different models can be compared to see which one it fits most closely. Such a scheme has several potential advantages because it does not require any pre-processing of the input text. We have performed extensive experiments on the use of PPM compression models for categorization using the standard Reuters-21578 dataset. We obtained some encouraging results on two-category situations, and the results on the general problem seem reasonably impressive---in one case outstanding. However, we find that PPM does not compete with the published state of the art in the use of machine learning for text categorization. It produces inferior results because it is insensitive to subtle differences between articles that belong to a category and those that do not. We do not believe our results are specific to PPM. If the occurrence of a single word determines whether an article belongs to a category or not (and it often does) any compression scheme will likely fail to classify the article correctly. Machine learning schemes fare better because they automatically eliminate irrelevant features and concentrate on the most discriminating ones.}, } @inProceedings{Frasconi01, author = {Paolo Frasconi and Giovanni Soda and Alessandro Vullo}, title = {Text Categorization for Multi-page Documents: A Hybrid Naive {Bayes HMM} Approach}, booktitle = {Proceedings of JCDL, 1st ACM-IEEE Joint Conference on Digital Libraries}, editor = {}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, year = {2001}, address = {Roanoke, US}, pages = {11--20}, url = {http://www.dsi.unifi.it/~paolo/ps/jcdl01-hmm-text.pdf}, abstract = {Text categorization is typically formulated as a concept learning problem where each instance is a single isolated document. In this paper we are interested in a more general formulation where documents are organized as page sequences, as naturally occurring in digital libraries of scanned books and magazines. We describe a method for classifying pages of sequential OCR text documents into one of several assigned categories and suggest that taking into account contextual information provided by the whole page sequence can significantly improve classification accuracy. The proposed architecture relies on hidden Markov models whose emissions are bag-of-words according to a multinomial word event model, as in the generative portion of the Naive Bayes classifier. Our results on a collection of scanned journals from the Making of America project confirm the importance of using whole page sequences. Empirical evaluation indicates that the error rate (as obtained by running a plain Naive Bayes classifier on isolated page) can be roughly reduced by half if contextual information is incorporated.}, } @article{Frasconi02, author = {Paolo Frasconi and Giovanni Soda and Alessandro Vullo}, title = {Text Categorization for Multi-page Documents: A Hybrid Naive {Bayes HMM} Approach}, journal = {Journal of Intelligent Information Systems}, year = {2002}, note = {Special Issue on Automated Text Categorization}, volume = {18}, number = {2/3}, pages = {195--217}, url = {http://www.wkap.nl/article.pdf?391247}, abstract = {In the traditional setting, text categorization is formulated as a concept learning problem where each instance is a single isolated document. However, this perspective is not appropriate in the case of many digital libraries that offer as contents scanned and optically read books or magazines. In this paper, we propose a more general formulation of text categorization, allowing documents to be organized as \textit{sequences} of pages. We introduce a novel hybrid system specifically designed for multi-page text documents. The architecture relies on hidden Markov models whose emissions are bag-of-words resulting from a multinomial word event model, as in the generative portion of the Naive Bayes classifier. The rationale behind our proposal is that taking into account contextual information provided by the whole page sequence can help disambiguation and improves single page classification accuracy. Our results on two datasets of scanned journals from the Making of America collection confirm the importance of using whole page sequences. The empirical evaluation indicates that the error rate (as obtained by running the Naive Bayes classifier on isolated pages) can be significantly reduced if contextual information is incorporated.}, } @inProceedings{Frommholz01, author = {Ingo Frommholz}, title = {Categorizing Web Documents in Hierarchical Catalogues}, booktitle = {Proceedings of ECIR-01, 23rd European Colloquium on Information Retrieval Research}, editor = {}, year = {2001}, address = {Darmstadt, DE}, publisher = {}, pages = {}, url = {http://ls6-www.informatik.uni-dortmund.de/bib/fulltext/ir/Frommholz:01a.pdf}, abstract = {Automatic categorization of web documents (e.g. HTML documents) denotes the task of automatically finding relevant categories for a (new) document which is to be inserted into a web catalogue like Yahoo!. There exist many approaches for performing this difficult task. Here, special kinds of web catalogues, those whose category scheme is hierarchically ordered, are regarded. A method for using the knowledge about the hierarchy to gain better categorization results is discussed. This method can be applied in a post-processing step and therefore be combined with other known (non-hierarchical) categorization approaches.}, } @inProceedings{Fuhr84, author = {Fuhr, Norbert and Knorz, Gerhard}, title = {Retrieval test evaluation of a rule-based automated indexing {(AIR/PHYS)}}, booktitle = {Proceedings of SIGIR-84, 7th ACM International Conference on Research and Development in Information Retrieval}, year = {1984}, publisher = {Cambridge University Press}, editor = {Van Rijsbergen, Cornelis J.}, pages = {391--408}, address = {Cambridge, UK}, url = {}, abstract = {}, } @inProceedings{Fuhr85, author = {Fuhr, Norbert}, title = {A probabilistic model of dictionary-based automatic indexing}, booktitle = {Proceedings of RIAO-85, 1st International Conference ``Recherche d'Information Assistee par Ordinateur''}, publisher = {}, editor = {}, address = {Grenoble, FR}, year = {1985}, pages = {207--216}, url = {}, abstract = {}, } @inProceedings{Fuhr91a, author = {Fuhr, Norbert and Hartmann, Stephan and Knorz, Gerhard and Lustig, Gerhard and Schwantner, Michael and Tzeras, Konstadinos}, title = {{AIR/X} -- a Rule-Based Multistage Indexing System for Large Subject Fields}, booktitle = {Proceedings of RIAO-91, 3rd International Conference ``Recherche d'Information Assistee par Ordinateur''}, publisher = {Elsevier Science Publishers, Amsterdam, NL}, editor = {Andr{\'e} Lichnerowicz}, address = {Barcelona, ES}, year = {1991}, pages = {606--623}, url = {http://www.darmstadt.gmd.de/~tzeras/FullPapers/gz/Fuhr-etal-91.ps.gz}, abstract = {AIR/X is a rule-based system for indexing with terms (descriptors) from a prescribed vocabulary. For this task, an indexing dictionary with rules for mapping terms from the text onto descriptors is required, which can be derived automatically from a set of manually indexed documents. Based on the Darmstadt Indexing Approach, the indexing task is divided into a description step and a decision step. First, terms (single words or phrases) are identified in the document text. With term-descriptor rules from the dictionary, descriptor indications are formed. The set of all indications from a document leading to the same descriptor is called a relevance description. A probabilistic classification procedure computes indexing weights for each relevance description. Since the whole system is rule-based, it can be adapted to different subject fields by appropriate modifications of the rule bases. A major application of AIR/X is the AIR/PHYS system developed for a large physics database. This application is described in more detail along with experimental results.}, } @inProceedings{Fuhr91b, author = {Norbert Fuhr and Ulrich Pfeifer}, title = {Combining Model-Oriented and Description-Oriented Approaches for Probabilistic Indexing}, booktitle = {Proceedings of SIGIR-91, 14th ACM International Conference on Research and Development in Information Retrieval}, editor = {Abraham Bookstein and Yves Chiaramella and Gerard Salton and Vijay V. Raghavan}, publisher = {ACM Press, New York, US}, address = {Chicago, US}, pages = {46--56}, year = {1991}, note = {An extended version appears as~\cite{Fuhr94}}, url = {http://www.acm.org/pubs/articles/proceedings/ir/122860/p46-fuhr/p46-fuhr.pdf}, abstract = {We distinguish model-oriented and description-oriented approaches in probabilistic information retrieval. The former refer to certain representations of documents and queries and use additional independence assumptions, whereas the latter map documents and queries onto feature vectors which form the input to certain classification procedures or regression methods. Description-oriented approaches are more flexible with respect to the underlying representations, but the definition of the feature vector is a heuristic step. In this paper, we combine a probabilistic model for the Darmstadt Indexing Approach with logistic regression. Here the probabilistic model forms a guideline for the definition of the feature vector. Experiments with the purely theoretical approach and with several heuristic variations show that heuristic assumptions may yield significant improvements.}, } @article{Fuhr94, author = {Norbert Fuhr and Ulrich Pfeifer}, title = {Probabilistic Information Retrieval as Combination of Abstraction Inductive Learning and Probabilistic Assumptions}, journal = {ACM Transactions on Information Systems}, year = {1994}, number = {1}, volume = {12}, pages = {92--115}, url = {http://ls6-www.informatik.uni-dortmund.de/bib/fulltext/ir/Fuhr_Pfeifer:94.ps.gz}, abstract = {We show that former approaches in probabilistic information retrieval are based on one or two of the three concepts abstraction, inductive learning and probabilistic assumptions, and we propose a new approach which combines all three concepts. This approach is illustrated for the case of indexing with a controlled vocabulary. For this purpose, we describe a new probabilistic model #rst, which is then combined with logistic regression, thus yielding a generalization of the original model. Experimental results for the pure theoretical model as well as for heuristic variants are given. Furthermore, linear and logistic regression are compared.}, } @article{Furnkranz02, author = {Johannes F{\"{u}}rnkranz}, title = {Hyperlink Ensembles: A Case Study in Hypertext Classification}, journal = {Information Fusion}, year = {2002}, number = {4}, volume = {3}, pages = {299--312}, url = {}, abstract = {In this paper, we introduce hyperlink ensembles, a novel type of ensemble classifier for classifying hypertext documents. Instead of using the text on a page for deriving features that can be used for training a classifier, we suggest to use portions of texts from all pages that point to the target page. A hyperlink ensemble is formed by obtaining one prediction for each hyperlink that points to a page. These individual predictions for each hyperlink are subsequently combined to a final prediction for the class of the target page. We explore four different ways of combining the individual predictions and four different techniques for identifying relevant text portions. The utility of our approach is demonstrated on a set of Web-pages that relate to Computer Science Departments.}, } @inProceedings{Furnkranz99, author = {Johannes F{\"{u}}rnkranz}, title = {Exploiting Structural Information for Text Classification on the WWW}, booktitle = {Proceedings of IDA-99, 3rd Symposium on Intelligent Data Analysis}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1642}, editor = {David J. Hand and Joost N. Kok and Michael R. Berthold}, address = {Amsterdam, NL}, year = {1999}, pages = {487--497}, url = {http://www.ai.univie.ac.at/~juffi/publications/ida-99.ps.gz}, abstract = {In this paper, we report on a set of experiments that explore the utility of making use of the structural information of WWW documents. Our working hypothesis is that it is often easier to classify a hypertext page using information provided on pages that point to it instead of using information that is provided on the page itself. We present experimental evidence that confirms this hypothesis on a set of Web pages that relate to computer science departments.} } @inProceedings{Gabrilovich:2004:NewsJunkie, author = "Gabrilovich, Evgeniy and Dumais, Susan and Horvitz, Eric", title = "Newsjunkie: {P}roviding Personalized Newsfeeds via Analysis of Information Novelty", booktitle = "Proceedings of the Thirteenth International World Wide Web Conference ({WWW2004})", year = 2004, pages = "482--490", month = "May", address = "New York, NY", publisher = "ACM Press", url = "http://www.cs.technion.ac.il/~gabr/papers/NewsJunkie.pdf", abstract = "We present a principled methodology for filtering news stories by formal measures of information novelty, and show how the techniques can be used to custom-tailor newsfeeds based on information that a user has already reviewed. We review methods for analyzing novelty and then describe NewsJunkie, a system that personalizes news for users by identifying the novelty of stories in the context of stories they have already reviewed. NewsJunkie employs novelty-analysis algorithms that represent articles as words and named entities. The algorithms analyze inter- and intra- document dynamics by considering how information evolves over time from article to article, as well as within individual articles. We review the results of a user study undertaken to gauge the value of the approach over legacy time-based review of newsfeeds, and also to compare the performance of alternate distance metrics that are used to estimate the dissimilarity between candidate new articles and sets of previously reviewed articles." } @inProceedings{Gabrilovich04, author = {Evgeniy Gabrilovich and Shaul Markovitch}, title = {Text categorization with many redundant features: using aggressive feature selection to make {SVM}s competitive with {C4.5}}, booktitle = {Proceedings of ICML-04, 21st International Conference on Machine Learning}, editor = {Carla E. Brodley}, year = {2004}, address = {Banff, CA}, pages = {}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.cs.technion.ac.il/~gabr/papers/fs-svm.pdf}, abstract = {Text categorization algorithms usually represent documents as bags of words and consequently have to deal with huge numbers of features. Most previous studies found that the majority of these features are relevant for classification, and that the performance of text categorization with support vector machines peaks when no feature selection is performed. We describe a class of text categorization problems that are characterized with many redundant features. Even though most of these features are relevant, the underlying concepts can be concisely captured using only a few features, while keeping all of them has substantially detrimental effect on categorization accuracy. We develop a novel measure that captures feature redundancy, and use it to analyze a large collection of datasets. We show that for problems plagued with numerous redundant features the performance of C4.5 is significantly superior to that of SVM, while aggressive feature selection allows SVM to beat C4.5 by a narrow margin.}, } @inProceedings{Galavotti00, author = {Luigi Galavotti and Fabrizio Sebastiani and Maria Simi}, title = {Experiments on the use of feature selection and negative evidence in automated text categorization}, booktitle = {Proceedings of ECDL-00, 4th European Conference on Research and Advanced Technology for Digital Libraries}, editor = {Jos{\'e} L. Borbinha and Thomas Baker}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1923}, year = {2000}, address = {Lisbon, PT}, pages = {59--68}, url = {http://www.math.unipd.it/~fabseb60/Publications/ECDL00.pdf}, abstract = {We tackle two different problems of {\em text categorization} (TC), namely feature selection and classifier induction. {\em Feature selection} (FS) refers to the activity of selecting, from the set of $r$ distinct features (i.e.\ words) occurring in the collection, the subset of $r'\ll r$ features that are most useful for compactly representing the meaning of the documents. We propose a novel FS technique, based on a simplified variant of the $\chi^2$ statistics. {\em Classifier induction} refers instead to the problem of automatically building a text classifier by learning from a set of documents pre-classified under the categories of interest. We propose a novel variant, based on the exploitation of negative evidence, of the well-known $k$-NN method. We report the results of systematic experimentation of these two methods performed on the standard {\sc Reuters-21578} benchmark.}, } @article{Gale93, author = {William A. Gale and Kenneth W. Church and David Yarowsky}, title = {A method for disambiguating word senses in a large corpus}, journal = {Computers and the Humanities}, year = {1993}, number = {5}, volume = {26}, pages = {415--439}, url = {http://www.research.att.com/~kwc/published_1993_sense.ps}, abstract = {}, } @inProceedings{Gao03, author = {Sheng Gao and Wen Wu and Chin-Hui Lee and Tat-Seng Chua}, title = {A maximal figure-of-merit learning approach to text categorization}, booktitle = {Proceedings of SIGIR-03, 26th ACM International Conference on Research and Development in Information Retrieval}, editor = {Jamie Callan and Gordon Cormack and Charles Clarke and David Hawking and Alan Smeaton}, publisher = {ACM Press, New York, US}, address = {Toronto, CA}, year = {2003}, pages = {174--181}, url = {http://doi.acm.org/10.1145/860435.860469}, abstract = {A novel maximal figure-of-merit (MFoM) learning approach to text categorization is proposed. Different from the conventional techniques, the proposed MFoM method attempts to integrate any performance metric of interest (e.g. accuracy, recall, precision, or F1 measure) into the design of any classifier. The corresponding classifier parameters are learned by optimizing an overall objective function of interest. To solve this highly nonlinear optimization problem, we use a generalized probabilistic descent algorithm. The MFoM learning framework is evaluated on the Reuters-21578 task with LSI-based feature extraction and a binary tree classifier. Experimental results indicate that the MFoM classifier gives improved F1 and enhanced robustness over the conventional one. It also outperforms the popular SVM method in micro-averaging F1. Other extensions to design discriminative multiple-category MFoM classifiers for application scenarios with new performance metrics could be envisioned too.}, } @inProceedings{Gaussier02, author = {{\'E}ric Gaussier and Cyril Goutte and Kris Popat and Francine Chen}, title = {A hierarchical model for clustering and categorising documents}, booktitle = {Proceedings of ECIR-02, 24th European Colloquium on Information Retrieval Research}, editor = {Fabio Crestani and Mark Girolami and Van Rijsbergen, Cornelis J.}, year = {2002}, address = {Glasgow, UK}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2291}, pages = {229--247}, url = {http://link.springer.de/link/service/series/0558/papers/2291/22910229.pdf}, abstract = {We propose a new hierarchical generative model for textual data, where words may be generated by topic specific distributions at any level in the hierarchy. This model is naturally well-suited to clustering documents in preset or automatically generated hierarchies, as well as categorising new documents in an existing hierarchy. Training algorithms are derived for both cases, and illustrated on real data by clustering news stories and categorising newsgroup messages. Finally, the generative model may be used to derive a Fisher kernel expressing similarity between documents.}, } @article{Gentili01, author = {G.L. Gentili and Mauro Marinilli and Alessandro Micarelli and Filippo Sciarrone}, title = {Text categorization in an intelligent agent for filtering information on the Web}, journal = {International Journal of Pattern Recognition and Artificial Intelligence}, pages = {527--549}, year = {2001}, number = {3}, volume = {15}, url = {http://www.worldscinet.com/journals/ijprai/15/preserved-docs/1503/S021800140100099X.pdf}, abstract = {This paper presents a text categorization system, capable of analyzing HTML/text documents collected from the Web. The system is a component of a more extensive intelligent agent for adaptive information filtering on the Web. It is based on a hybrid case-based architecture, where two multilayer perceptrons are integrated into a case-based reasoner. An empirical evaluation of the system was performed by means of a confidence interval technique. The experimental results obtained are encouraging and support the choice of a hybrid case-based approach to text categorization.}, } @inProceedings{Geutner93, author = {Petra Geutner and Uli Bodenhausen and Alex Waibel}, title = {Flexibility Through Incremental Learning: Neural Networks for Text Categorization}, booktitle = {Proceedings of WCNN-93, World Congress on Neural Networks}, publisher = {}, editor = {}, year = {1993}, address = {Portland, US}, pages = {24--27}, url = {http://werner.ira.uka.de/papers/speech/1993/WCNN_93_petra_geutner.ps.gz}, abstract = {In this paper we show an adaptive incremental learning algorithm that learns interactively to classify text messages (here: emails) into categories without the need for lengthy batch training runs. The algorithm was evaluated on a large database of email messages that fall into five subjective categories. As control experiment best human categorization performance was established at 79.4\% for this task. The best of all connectionist architectures presented here achieves near human performance (79.1\%). This architecture acquires its language model and dictionary adaptively and hence avoids handcoding of either. The learning algorithm combines an adaptive phase which instantly updates dictionary and weights during interaction and a tuning phase which fine tunes for performance using previously seen data. Such systems can be deployed in various applications where instantaneous interactive learning is necessary such as on-line email or news categorization, text summarization and information filtering in general.}, } @inProceedings{Ghani00, author = {Rayid Ghani}, title = {Using error-correcting codes for text classification}, booktitle = {Proceedings of ICML-00, 17th International Conference on Machine Learning}, editor = {Pat Langley}, year = {2000}, address = {Stanford, US}, pages = {303--310}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.cs.cmu.edu/~rayid/mypapers/ecoc-icml.ps}, abstract = {This paper explores in detail the use of Error Correcting Output Coding (ECOC) for learning text classifiers. We show that the accuracy of a Naive Bayes Classifier over text classification tasks can be significantly improved by taking advantage of the error-correcting properties of the code. We also explore the use of different kinds of codes, namely Error-Correcting Codes, Random Codes, and Domain and Data-specific codes and give experimental results for each of them. The ECOC method scales well to large data sets with a large number of classes. Experiments on a real-world data set show a reduction in classification error by up to 66\% over the traditional Naive Bayes Classifier. We also compare our empirical results to semi-theoretical results and find that the two closely agree.}, } @inProceedings{Ghani01, author = {Rayid Ghani and Se{\'{a}}n Slattery and Yiming Yang}, title = {Hypertext Categorization using Hyperlink Patterns and Meta Data}, booktitle = {Proceedings of ICML-01, 18th International Conference on Machine Learning}, editor = {Carla Brodley and Andrea Danyluk}, address = {Williams College, US}, year = {2001}, pages = {178--185}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.cs.cmu.edu/~yiming/papers.yy/hypertext-icml01.ps.gz}, abstract = {Hypertext poses new text classification research challenges as hyperlinks, content of linked documents, and meta data about related web sites all provide richer sources of information for hypertext classification that are not available in traditional text classification. We investigate the use of such information for representing web sites, and the effectiveness of different classifiers (Naive Bayes, Nearest Neighbor, and {\sc Foil}) in exploiting those representations. We find that using words in web pages alone often yields suboptimal performance of classifiers, compared to exploiting additional sources of information beyond document content. On the other hand, we also observe that linked pages can be more harmful than helpful when the linked neighborhoods are highly ``noisy'' and that links have to be used in a careful manner. More importantly, our investigation suggests that meta data which is often available, or can be acquired using Information Extraction techniques, can be extremely useful for improving classification accuracy. Finally, the relative performance of the different classifiers being tested gives us insights into the strengths and limitations of our algorithms for hypertext classification.}, } @inProceedings{Ghani01a, author = {Rayid Ghani}, title = {Combining Labeled and Unlabeled data for Text Classification with a Large Number of Categories}, booktitle = {Proceedings of the IEEE International Conference on Data Mining}, editor = {Nick Cercone and Tsau Young Lin and Xindong Wu}, address = {San Jose, US}, year = {2001}, pages = {597--598}, publisher = {IEEE Computer Society, Los Alamitos, US}, url = {http://www.cs.cmu.edu/~rayid/mypapers/icdm01.ps}, abstract = {We develop a framework to incorporate unlabeled data in the Error-Correcting Output Coding (ECOC) setup by decomposing multiclass problems into multiple binary problems and then use Co-Training to learn the individual binary classification problems. We show that our method is especially useful for classification tasks involving a large number of categories where Co-training doesn't perform very well by itself and when combined with ECOC, outperforms several other algorithms that combine labeled and unlabeled data for text classification in terms of accuracy, precision-recall tradeoff, and efficiency.}, } @inProceedings{Ghani02, author = {Rayid Ghani}, title = {Combining Labeled and Unlabeled Data for MultiClass Text Categorization}, booktitle = {Proceedings of ICML-02, 19th International Conference on Machine Learning}, editor = {}, year = {2002}, address = {Sydney, AU}, pages = {}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.accenture.com/xdoc/en/services/technology/publications/Ghani-ICML02.pdf}, abstract = {Supervised learning techniques for text classification often require a large number of labeled examples to learn accurately. One way to reduce the amount of labeled data required is to develop algorithms that can learn effectively from a small number of labeled examples augmented with a large number of unlabeled examples. Current text learning techniques for combining labeled and unlabeled, such as EM and Co-Training, are mostly applicable for classification tasks with a small number of classes and do not scale up well for large multiclass problems. In this paper, we develop a framework to incorporate unlabeled data in the Error-Correcting Output Coding (ECOC) setup by first decomposing multiclass problems into multiple binary problems and then using Co-Training to learn the individual binary classification problems. We show that our method is especially useful for text classification tasks involving a large number of categories and outperforms other semi-supervised learning techniques such as EM and Co-Training. In addition to being highly accurate, this method utilizes the hamming distance from ECOC to provide high-precision results. We also present results with algorithms other than co-training in this framework and show that co-training is uniquely suited to work well within ECOC.}, } @inProceedings{Giorgetti03, author = {Daniela Giorgetti and Fabrizio Sebastiani}, title = {Multiclass Text Categorization for Automated Survey Coding}, year = {2003}, address = {Melbourne, US}, booktitle = {Proceedings of SAC-03, 18th ACM Symposium on Applied Computing}, publisher = {ACM Press, New York, US}, pages = {798--802}, url = {http://www.math.unipd.it/~fabseb60/Publications/SAC03a.pdf}, abstract = {\emph{Survey coding} is the task of assigning a symbolic code from a predefined set of such codes to the answer given in response to an open-ended question in a questionnaire (aka \emph{survey}). We formulate the problem of automated survey coding as a \emph{text categorization} problem, i.e.\ as the problem of learning, by means of supervised machine learning techniques, a model of the association between answers and codes from a training set of pre-coded answers, and applying the resulting model to the classification of new answers. In this paper we experiment with two different learning techniques, one based on na\"{\i}ve Bayesian classification and the other one based on multiclass support vector machines, and test the resulting framework on a corpus of social surveys. The results we have obtained significantly outperform the results achieved by previous automated survey coding approaches.}, } @article{Giorgetti03a, author = {Daniela Giorgetti and Fabrizio Sebastiani}, title = {Automating Survey Coding by Multiclass Text Categorization Techniques}, journal = {Journal of the American Society for Information Science and Technology}, year = {2003}, volume = {54}, number = {12}, pages = {1269--1277}, url = {http://www.math.unipd.it/~fabseb60/Publications/JASIST03.pdf}, abstract = {\emph{Survey coding} is the task of assigning a symbolic code from a predefined set of such codes to the answer given in response to an open-ended question in a questionnaire (aka \emph{survey}). This task is usually carried out in order to group respondents according to a predefined scheme based on their answers. Survey coding has several applications, especially in the social sciences, ranging from the simple classification of respondents to the extraction of statistics on political opinions, health and lifestyle habits, customer satisfaction, brand fidelity, and patient satisfaction. Survey coding is a difficult task, since the code that should be attributed to a respondent based on the answer she has given is a matter of subjective judgment, and thus requires expertise. It is thus unsurprising that this task has traditionally been performed manually, by trained coders. Some attempts have been made at automating this task, most of them based on detecting the similarity between the answer and textual descriptions of the meanings of the candidate codes. We take a radically new stand, and formulate the problem of automated survey coding as a \emph{text categorization} problem, i.e.\ as the problem of learning, by means of supervised machine learning techniques, a model of the association between answers and codes from a training set of pre-coded answers, and applying the resulting model to the classification of new answers. In this paper we experiment with two different learning techniques, one based on na\"{\i}ve Bayesian classification and the other one based on multiclass support vector machines, and test the resulting framework on a corpus of social surveys. The results we have obtained significantly outperform the results achieved by previous automated survey coding approaches.}, } @inProceedings{Glover02, author = {Eric J. Glover and Kostas Tsioutsiouliklis and Steve Lawrence and David M. Pennock and Gary W. Flake}, title = {Using Web structure for classifying and describing Web pages}, booktitle = {Proceedings of WWW-02, International Conference on the World Wide Web}, address = {Honolulu, US}, year = {2002}, pages = {562--569}, publisher = {ACM Press, New York, US}, url = {http://www.cs.princeton.edu/~kt/www02.ps}, abstract = {The structure of the web is increasingly being used to improve organization, search, and analysis of information on the web. For example, Google uses the text in citing documents (documents that link to the target document) for search. We analyze the relative utility of document text, and the text in citing documents near the citation, for classification and description. Results show that the text in citing documents, when available, often has greater discriminative and descriptive power than the text in the target document itself. The combination of evidence from a document and citing documents can improve on either information source alone. Moreover, by ranking words and phrases in the citing documents according to expected entropy loss, we are able to accurately name clusters of web pages, even with very few positive examples. Our results confirm, quantify, and extend previous research using web structure in these areas, introducing new methods for classification and description of pages.}, } @inProceedings{Godbole04, author = {Shantanu Godbole and Abhay Harpale and Sunita Sarawagi and Soumen Chakrabarti}, title = {Document classification through interactive supervision of document and term labels}, booktitle = {Proceedings of PKDD-04, 8th European Conference on Principles of Data Mining and Knowledge Discovery}, editor = {Jean-Fran{\c{c}}ois Boulicaut and Floriana Esposito and Fosca Giannotti and Dino Pedreschi}, address = {Pisa, IT}, pages = {185--196}, year = {2004}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 3202}, url = {}, abstract = {}, } @inProceedings{Goevert99, author = {Norbert G{\"{o}}vert and Mounia Lalmas and Norbert Fuhr}, title = {A probabilistic description-oriented approach for categorising Web documents}, booktitle = {Proceedings of CIKM-99, 8th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {1999}, address = {Kansas City, US}, pages = {475--482}, url = {http://ls6-www.informatik.uni-dortmund.de/ir/publications/1999/Goevert_etal:99.html}, abstract = {The automatic categorisation of web documents is becoming crucial for organising the huge amount of information available in the Internet. We are facing a new challenge due to the fact that web documents have a rich structure and are highly heterogeneous. Two ways to respond to this challenge are (1) using a representation of the content of web documents that captures these two characteristics and (2) using more effective classifiers. Our categorisation approach is based on a probabilistic description-oriented representation of web documents, and a probabilistic interpretation of the k-nearest neighbour classifier. With the former, we provide an enhanced document representation that incorporates the structural and heterogeneous nature of web documents. With the latter, we provide a theoretical sound justification for the various parameters of the k-nearest neighbour classifier. Experimental results show that (1) using an enhanced representation of web documents is crucial for an effective categorisation of web documents, and (2) a theoretical interpretation of the k-nearest neighbour classifier gives us improvement over the standard k-nearest neighbour classifier.}, } @inProceedings{Goldberg95, author = {Goldberg, Jeffrey L.}, title = {CDM: an approach to learning in text categorization}, booktitle = {Proceedings of ICTAI-95, 7th International Conference on Tools with Artificial Intelligence}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, editor = {}, address = {Herndon, US}, year = {1995}, pages = {258--265}, url = {}, note = {An extended version appears as~\cite{Goldberg96}}, abstract = {The category discrimination method (CDM) is a new learning algorithm designed for text categorization. The motivation is that there are statistical problems associated with natural language text when it is applied as input to existing machine learning algorithms (too much noise, too many features, skewed distribution). The bases of the CDM are research results about the way that humans learn categories and concepts vis-a-vis contrasting concepts. The essential formula is cue validity borrowed from cognitive psychology, and used to select from all possible single word-based features the `best` predictors of a given category. The hypothesis that CDM`s performance exceeds two non-domain specific algorithms, Bayesian classification and decision tree learners, is empirically tested.}, } @article{Goldberg96, author = {Goldberg, Jeffrey L.}, title = {CDM: an approach to learning in text categorization}, journal = {International Journal on Artificial Intelligence Tools}, year = {1996}, number = {1/2}, volume = {5}, pages = {229--253}, url = {}, abstract = {The Category Discrimination Method (CDM) is a new machine learning algorithm designed specifically for text categorization. The motivation is that there are statistical problems associated with natural language text when it is applied as input to existing machine learning algorithms (too much noise, too many features, skewed distribution). The bases of the CDM are research results about the way that humans learn categories and concepts vis-a-vis contrasting concepts. The essential formula is cue validity borrowed from cognitive psychology, and used to select from all possible single word based features, the best predictors of a given category. The hypothesis that CDM's performance will exceed two non domain specific algorithms, Bayesian classification and decision tree learners, is empirically tested.}, } @inProceedings{Gomez02, author = {G{\'o}mez-Hidalgo, Jos{\'e} M. and De Buenaga Rodr{\'{\i}}guez, Jos{\'e} M. and Ureña L{\'o}pez, Luis A. and Mart{\'{\i}}n Valdivia, Maria T. and Garc{\'{\i}}a Vega, Manuel}, title = {Integrating Lexical Knowledge in Learning-Based Text Categorization}, booktitle = {Proceedings of JADT-02, 6th International Conference on the Statistical Analysis of Textual Data}, publisher = {}, editor = {}, address = {St-Malo, FR}, pages = {}, year = {2002}, url = {http://www.cavi.univ-paris3.fr/lexicometrica/jadt/jadt2002/PDF-2002/gomez_debuenaga_urena_martin_garcia.pdf}, abstract = {Automatic Text Categorization (ATC) is an important task in the field of Information Access. The prevailing approach to ATC is making use of a collection of prelabeled texts for the induction of a document classifier through learning methods. With the increasing availability of lexical resources in electronic form (including Lexical Databases (LDBs), Machine Readable Dictionaries, etc.), there is an interesting opportunity for the integration of them in learning-based ATC. In this paper, we present an approach to the integration of lexical knowledge extracted from the LDB WordNet in learning-based ATC, based on Stacked Generalization (SG). The method we suggest is based on combining the lexical knowledge extracted from the LDB interpreted as a classifier with a learning-based classifier, through SG. We have performed experiments which results show that the ideas we describe are promising and deserve further investigation.}, } @inProceedings{Gomez02a, author = {G{\'o}mez-Hidalgo, Jos{\'e} M.}, title = {Evaluating Cost-Sensitive Unsolicited Bulk Email Categorization}, booktitle = {Proceedings of SAC-02, 17th ACM Symposium on Applied Computing}, editor = {}, address = {Madrid, ES}, pages = {615--620}, year = {2002}, url = {http://doi.acm.org/10.1145/508791.508911}, abstract = {In the recent years, Unsolicited Bulk Email has became an increasingly important problem, with a big economic impact. In this paper, we discuss cost-sensitive Text Categorization methods for UBE filtering. In concrete, we have evaluated a range of Machine Learning methods for the task (C4.5, Naive Bayes, PART, Support Vector Machines and Rocchio), made cost sensitive through several methods (Threshold Optimization, Instance Weighting, and Meta-Cost). We have used the Receiver Operating Characteristic Convex Hull method for the evaluation, that best suits classification problems in which target conditions are not known, as it is the case. Our results do not show a dominant algorithm nor method for making algorithms cost-sensitive, but are the best reported on the test collection used, and approach real-world hand-crafted classifiers accuracy.}, } @inProceedings{Goodman90, author = {Marc Goodman}, title = {{\sc Prism}: a case-based telex classifier}, booktitle = {Proceedings of IAAI-90, 2nd Conference on Innovative Applications of Artificial Intelligence}, publisher = {AAAI Press, Menlo Park, US}, editor = {Alain Rappaport and Reid Smith}, year = {1990}, address = {}, pages = {25--37}, url = {}, abstract = {}, } @article{Gray71, author = {W. A. Gray and A. J. Harley}, title = {Computer-assisted indexing}, journal = {Information Storage and Retrieval}, year = {1971}, volume = {7}, number = {4}, pages = {167--174}, url = {}, abstract = {}, } @inProceedings{Guo04, author = {Gongde Guo and Hui Wang and David A. Bell and Yaxin Bi and Kieran Greer}, title = {An kNN Model-Based Approach and Its Application in Text Categorization}, booktitle = {Proceedings of CICLING-04, 5th International Conference on Computational Linguistics and Intelligent Text Processing}, year = {2004}, editor = {Alexander F. Gelbukh}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Seoul, KO}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2945}, pages = {559--570}, url = {}, abstract = {}, } @inProceedings{Guthrie94, author = {Louise Guthrie and Elbert Walker and Joe A. Guthrie}, title = {Document classification by machine: theory and practice}, booktitle = {Proceedings of COLING-94, 15th International Conference on Computational Linguistics}, publisher = {}, editor = {}, address = {Kyoto, JP}, year = {1994}, pages = {1059--1063}, url = {}, abstract = {}, } @inCollection{Guthrie99, author = {Louise Guthrie and Joe A. Guthrie and James Leistensnider}, title = {Document classification and routing}, booktitle = {Natural language information retrieval}, editor = {Tomek Strzalkowski}, year = {1999}, pages = {289--310}, publisher = {Kluwer Academic Publishers}, address = {Dordrecht, NL}, url = {}, abstract = {}, } @inProceedings{Hadjarian01, author = {Ali Hadjarian and Jerzy Bala and Peter Pachowicz}, title = {Text Categorization through Multistrategy Learning and Visualization}, booktitle = {Proceedings of CICLING-01, 2nd International Conference on Computational Linguistics and Intelligent Text Processing}, year = {2001}, editor = {Alexander Gelbukh}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Mexico City, ME}, note = {Published in the ``Lecture Notes for Computer Science'' series, number 2004}, pages = {423--436}, url = {http://link.springer.de/link/service/series/0558/papers/2004/20040437.pdf}, abstract = {This paper introduces a multistrategy learning approach to the categorization of text documents. The approach benefits from two existing, and in our view complimentary, sets of categorization techniques: those based on Rocchio's algorithm and those belonging to the rule learning class of machine learning algorithms. Visualization is used for the presentation of the output of learning}, } @inProceedings{Hamill78, author = {Hamill, Karen A. and Zamora, Antonio}, title = {An automatic document classification system using pattern recognition techniques}, booktitle = {Proceedings of ASIS-78, 41st Annual Meeting of the American Society for Information Science}, publisher = {American Society for Information Science, Washington, US}, editor = {Everett H. Brenner}, year = {1978}, address = {New York, US}, pages = {152--155}, url = {}, abstract = {}, } @article{Hamill80, author = {Hamill, Karen A. and Zamora, Antonio}, title = {The Use of titles for Automatic Document Classification}, journal = {Journal of the American Society for Information Science}, year = {1980}, number = {6}, pages = {396--402}, volume = {33}, url = {}, abstract = {}, } @inProceedings{Han01, author = {Eui-Hong Han and George Karypis and Vipin Kumar}, title = {Text Categorization Using Weight-Adjusted $k$-Nearest Neighbor Classification}, booktitle = {Proceedings of PAKDD-01, 5th Pacific-Asia Conferenece on Knowledge Discovery and Data Mining}, editor = {David Cheung and Qing Li and Graham Williams}, year = {2001}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Hong Kong, CN}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2035}, pages = {53--65}, url = {http://link.springer.de/link/service/series/0558/papers/2035/20350053.pdf}, abstract = {Text categorization presents unique challenges due to the large number of attributes present in the data set, large number of training samples, attribute dependency, and multi-modality of categories. Existing classification techniques have limited applicability in the data sets of these natures. In this paper, we present a Weight Adjusted k-Nearest Neighbor (WAKNN) classification that learns feature weights based on a greedy hill climbing technique. We also present two performance optimizations of WAKNN that improve the computational performance by a few orders of magnitude, but do not compromise on the classification quality. We experimentally evaluated WAKNN on 52 document data sets from a variety of domains and compared its performance against several classification algorithms, such as C4.5, RIPPER, Naive-Bayesian, PEBLS and VSM. Experimental results on these data sets confirm that WAKNN consistently outperforms other existing classification algorithms.}, } @article{Hanauer96, author = {David Hanauer}, title = {Integration of phonetic and graphic features in poetic text categorization judgements}, journal = {Poetics}, year = {1996}, volume = {23}, number = {5}, pages = {363--380}, url = {}, abstract = {The experiments reported in this paper deal with the relationship between specific formal textual features, i.e. graphic and phonetic information, and the reader's literary educational background in the categorization of poetic texts. In two experiments, the research method of Information Integration Theory was employed in order to test two hypotheses relating to the radical conventionalist and traditional positions on the role of specific formal textual features in the categorization of poetic texts. Twenty subjects from expert or novice literary reading experience backgrounds were, in two experiments, required to rate two parallel sets of graphically and phonetically manipulated poems. The results reveal that subjects are sensitive to the manipulations of graphic and phonetic information and use the same additive information integration rule in making poetic text categorization judgements. The expert literary readers were found to assign significantly higher ratings to all versions of the manipulated poems than the novice readers.}, } @inProceedings{Hayes88, author = {Philip J. Hayes and Laura E. Knecht and Monica J. Cellio}, title = {A news story categorization system}, booktitle = {Proceedings of ANLP-88, 2nd Conference on Applied Natural Language Processing}, publisher = {Association for Computational Linguistics, Morristown, US}, address = {Austin, US}, editor = {}, year = {1988}, pages = {9--17}, url = {}, note = {Reprinted in Karen Sparck Jones and Peter Willett (eds.), ``Readings in Information Retrieval'', Morgan Kaufmann, San Francisco, US, 1997, pp.\ 518--526.}, abstract = {The article describes a pilot version of a commercial application of natural language processing techniques to the problem of categorizing new stories into broad topic categories. The system does not perform a complete semantic or syntactic analyses of the input stories. Its categorizations are dependent on fragmentary recognition using pattern-matching techniques. The fragments it looks for are determined by a set of knowledge-based rules. The accuracy of the system is only slightly lower than that of human categorizers.}, } @inProceedings{Hayes90, author = {Philip J. Hayes and Steven P. Weinstein}, title = {{\sc Construe/Tis}: a system for content-based indexing of a database of news stories}, booktitle = {Proceedings of IAAI-90, 2nd Conference on Innovative Applications of Artificial Intelligence}, publisher = {AAAI Press, Menlo Park, US}, editor = {Alain Rappaport and Reid Smith}, year = {1990}, pages = {49--66}, url = {}, abstract = {}, address = {Boston, US}, } @inProceedings{Hayes90a, author = {Philip J. Hayes and Peggy M. Andersen and Irene B. Nirenburg and Linda M. Schmandt}, title = {{\sc Tcs}: a shell for content-based text categorization}, booktitle = {Proceedings of CAIA-90, 6th IEEE Conference on Artificial Intelligence Applications}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, editor = {}, year = {1990}, address = {Santa Barbara, US}, pages = {320--326}, url = {}, abstract = {The kind of application that the text categorization shell, TCS, can produce is characterized. Many of its applications have great commercial value. The design goals for TCS are discussed, and other approaches to text categorization in the light of these goals are examined. The TCS and how it meets its design goals are described, and examples of applications built with TCS are given. A text-categorization application developed with TCS consists of the TCS run-time system and a rule base. The rule base defines what categories the application can assign to texts and contains rules that make the categorization decisions for particular texts. The data-driven nature of TCS allows it is to satisfy fully the requirements of ease of application development, portability to other applications and maintainability.}, } @article{He03, author = {Ji He and Ah-Hwee Tan and Chew-Lim Tan}, title = {On Machine Learning Methods for Chinese Document Categorization}, journal = {Applied Intelligence}, year = {2003}, volume = {18}, number = {3}, pages = {311--322}, url = {http://www.kluweronline.com/issn/0924-669X}, abstract = {This paper reports our comparative evaluation of three machine learning methods, namely k Nearest Neighbor (kNN), SupportVector Machines (SVM), and Adaptive Resonance Associative Map (ARAM) for Chinese document categorization. Based on two Chinese corpora, a series of controlled experiments evaluated their learning capabilities and efficiency in mining text classification knowledge. Benchmark experiments showed that their predictive performance were roughly comparable, especially on clean and well organized data sets. While kNN and ARAM yield better performances than SVM on small and clean data sets, SVM and ARAM significantly outperformed kNN on noisy data. Comparing efficiency, kNN was notably more costly in terms of time and memory than the other two methods. SVM is highly efficient in learning from well organized samples of moderate size, although on relatively large and noisy data the efficiency of SVM and ARAM are comparable.}, } @article{Heaps73, author = {H.S. Heaps}, title = {A theory of relevance for automatic document classification}, year = {1973}, journal = {Information and Control}, volume = {22}, number = {3}, pages = {268-278}, url = {}, abstract = {}, } @inProceedings{Hearst91, author = {Marti A. Hearst}, title = {Noun homograph disambiguation using local context in large corpora}, booktitle = {Proceedings of the 7th Annual Conference of the University of Waterloo Centre for the New Oxford English Dictionary}, publisher = {}, editor = {}, year = {1991}, pages = {1--22}, address = {Oxford, UK}, url = {ftp://parcftp.xerox.com/pub/hearst/oed91.ps.gz}, abstract = {This paper describes an accurate, relatively inexpensive method for the disambiguation of noun homographs using large text corpora. The algorithm checks the context surrounding the target noun against that of previously observed instances and chooses the sense for which the most evidence is found, where evidence consists of a set of orthographic, syntactic, and lexical features. Because the sense distinctions made are coarse, the disambiguation can be accomplished without the expense of knowledge bases or inference mechanisms. An implementation of the algorithm is described which, starting with a small set of hand-labeled instances, improves its results automatically via unsupervised training. The approach is compared to other attempts at homograph disambiguation using both machine readable dictionaries and unrestricted text and the use of training instances is determined to be a crucial difference.}, } @proceedings{Hearst96a, editor = {Marti A. Hearst and Haym Hirsh}, title = {Machine Learning in Information Access. Papers from the 1996 AAAI Spring Symposium}, institution = {Americal Association for Artificial Intelligence}, address = {Stanford, US}, year = {1996}, note = {Available as Technical Report SS-96-05}, url = {}, abstract = {}, } @inProceedings{Hersh94, author = {William Hersh and Christopher Buckley and T.J. Leone and David Hickman}, title = {{{\sc Ohsumed}}: an interactive retrieval evaluation and new large text collection for research}, booktitle = {Proceedings of SIGIR-94, 17th ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and Van Rijsbergen, Cornelis J.}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Dublin, IE}, pages = {192--201}, year = {1994}, url = {http://www.acm.org/pubs/articles/proceedings/ir/188490/p192-hersh/p192-hersh.pdf}, abstract = {A series of information retrieval experiments was carried out with a computer installed in a medical practice setting for relatively inexperienced physician end-users. Using a commercial MEDLINE product based on the vector space model, these physicians searched just as effectively as more experienced searchers using Boolean searching. The results of this experiment were subsequently used to create a new large medical test collection, which was used in experiments with the SMART retrieval system to obtain baseline performance data as well as compare SMART with the other searchers.}, } @inProceedings{Hoashi00, author = {Keiichiro Hoashi and Kazunori Matsumoto and Naomi Inoue and Kazuo Hashimoto}, title = {Document filtering methods using non-relevant information profile}, booktitle = {Proceedings of SIGIR-00, 23rd ACM International Conference on Research and Development in Information Retrieval}, editor = {Nicholas J. Belkin and Peter Ingwersen and Mun-Kew Leong}, publisher = {ACM Press, New York, US}, address = {Athens, GR}, year = {2000}, pages = {176--183}, url = {http://www.acm.org/pubs/articles/proceedings/ir/345508/p176-hoashi/p176-hoashi.pdf}, abstract = {Document filtering is a task to retrieve documents relevant to a user's profile from a flow of documents. Generally, filtering systems calculate the similarity between the profile and each incoming document, and retrieve documents with similarity higher than a threshold. However, many systems set a relatively high threshold to reduce retrieval of non-relevant documents, which results in the ignorance of many relevant documents. In this paper, we propose the use of a non-relevant information profile to reduce the mistaken retrieval of non-relevant documents. Results from experiments show that this filter has successfully rejected a sufficient number of non-relevant documents, resulting in an improvement of filtering performance.}, } @inProceedings{Hoch94, author = {Rainer Hoch}, title = {Using IR techniques for text classification in document analysis}, booktitle = {Proceedings of SIGIR-94, 17th ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and Van Rijsbergen, Cornelis J.}, publisher = {Springer Verlag, Heidelberg, DE}, year = {1994}, address = {Dublin, IE}, pages = {31--40}, url = {http://www.acm.org/pubs/articles/proceedings/ir/188490/p31-hoch/p31-hoch.pdf}, abstract = {This paper presents the INFOCLAS system applying statistical methods of information retrieval for the classification of German business letters into corresponding message types such as order, offer, enclosure, etc. INFOCLAS is a first step towards the understanding of documents proceeding to a classification-driven extraction of information. The system is composed of two main modules: the central indexer (extraction and weighting of indexing terms) and the classifier (classification of business letters into given types). The system employs several knowledge sources including a letter database, word frequency statistics for German, lists of message type specific words, morphological knowledge as well as the underlying document structure. As output, the system evaluates a set of weighted hypotheses about the type of the actual letter. Classification of documents allow the automatic distribution or archiving of letters and is also an excellent starting point for higher-level document analysis.}, } @article{Hoyle73, author = {W.G. Hoyle}, title = {Automatic indexing and generation of classification by algorithm}, journal = {Information Storage and Retrieval}, year = {1973}, volume = {9}, number = {4}, pages = {233--242}, url = {}, abstract = {A system of automatic indexing based on Bayes' theorem is described briefly. In assigning 124 documents to 9 categories, there were 97 cases of agreement with professional indexers. Using a collection factor, based on 87 per cent human consistency from other courses, the computer appears then to index with 90 per cent accuracy in this case. The technique is then used with two randomized sample document groups drawn from nine categories. Each group in turn is used as the basis for indexing the other. The computer knows only the number of categories. After 8 cycles the computer is found to have formed 9 groups consisting of about 50 per cent of documents that were also lumped together by professional indexers on the basis of subject content. A new measure of performance is proposed and some other applications of the technique indicated.}, } @inProceedings{Hsu99, author = {Wen-Lin Hsu and Sheau-Dong Lang}, title = {Classification algorithms for NETNEWS articles}, booktitle = {Proceedings of CIKM-99, 8th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {1999}, address = {Kansas City, US}, pages = {114--121}, url = {http://www.acm.org/pubs/articles/proceedings/cikm/319950/p114-hsu/p114-hsu.pdf}, abstract = {We propose several algorithms using the vector space model to classify the news articles posted on the NETNEWS according to the newsgroup categories. The baseline method combines the terms of all the articles of each newsgroup in the training set to represent the newsgroups as single vectors. After training, the incoming news articles are classified based on their similarity to the existing newsgroup categories. We propose to use the following techniques to improve the classification performance of the baseline method: (1) use routing (classification) accuracy and the similarity values to refine the training set; (2) update the underlying term structures periodically during testing; and (3) apply k-means clustering to partition the newsgroup articles and represent each newsgroup by k vectors. Our test collection consists of the real news articles and the 519 subnewsgroups under the REC newsgroup of NETNEWS in a period of 3 months. Our experimental results demonstrate that the technique of refining the training set reduces from one-third to two-thirds of the storage. The technique of periodical updates improves the routing accuracy ranging from 20\% to 100\% but incurs runtime overhead. Finally, representing each newsgroup by k vectors (with k = 2 or 3) using clustering yields the most significant improvement in routing accuracy, ranging from 60\% to lOO\%, while causing only slightly higher storage requirements.}, } @inProceedings{Hsu99a, author = {Wen-Lin Hsu and Sheau-Dong Lang}, title = {Feature Reduction and Database Maintenance in NETNEWS Classification}, booktitle = {Proceedings of IDEAS-99, 1999 International Database Engineering and Applications Symposium}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, editor = {}, year = {1999}, address = {Montreal, CA}, pages = {137--144}, url = {http://dlib.computer.org/conferen/ideas/0265/pdf/02650137.pdf}, abstract = {We propose a statistical feature-reduction technique to filter out the most ambiguous articles in the training data for categorizing the NETNEWS articles. We also incorporate a batch updating scheme to periodically do maintenance on the term structures of the news database after training. The baseline method combines the terms of all the articles of each newsgroup in the training set to represent the newsgroups as single vectors. After training, the incoming news articles are classified based on their similarity to the existing newsgroup categories. Our implementation uses an inverted file to store the trained term structures of each newsgroup, and uses a list similar to the inverted file to buffer the newly arrival articles, for efficient routing and updating purposes. Our experimental results using real NETNEWS articles and newsgroups demonstrate (1) applying feature reduction to the training set improves the routing accuracy, efficiency, and database storage; (2) updating improves the routing accuracy; and (3) the batch technique improves the efficiency of the updating operation.}, } @inProceedings{Huffman94, author = {Stephen Huffman and Marc Damashek}, title = {Acquaintance: A Novel Vector-Space N-Gram Technique for Document Categorization}, booktitle = {Proceedings of TREC-3, 3rd Text Retrieval Conference}, publisher = {National Institute of Standards and Technology, Gaithersburg, US}, editor = {Donna K. Harman}, year = {1994}, address = {Gaithersburg, US}, pages = {305--310}, url = {}, abstract = {Acquaintance is the name of a novel vector-space n-gram technique for categorizing documents. The technique is completely language-independent, highly garble-resistant, and computationally simple. An unoptimized version of the algorithm was used to process the TREC database in a very short time.}, } @inProceedings{Huffman95, author = {Stephen Huffman}, title = {Acquaintance: Language-Independent Document Categorization by N-Grams}, booktitle = {Proceedings of TREC-4, 4th Text Retrieval Conference}, publisher = {National Institute of Standards and Technology, Gaithersburg, US}, editor = {Donna K. Harman and Ellen M. Voorhees}, year = {1995}, address = {Gaithersburg, US}, pages = {359--371}, url = {http://trec.nist.gov/pubs/trec4/papers/nsa.ps.gz}, abstract = {Acquaintance is the name of a novel vector-space n-gram for categorizing documents. The technique is completely language-independent, highly garble-resistant, and computationally simple. An unoptimized version of the algorithm was used to process the TREC database in a very short time. The TREC-3 conference provided the first public demonstration and evaluation of this new technique, and TREC-4 provided an opportunity to test its usefulness on several types of text retrieval tasks.}, } @inProceedings{Hull94, author = {Hull, David A.}, title = {Improving text retrieval for the routing problem using latent semantic indexing}, booktitle = {Proceedings of the 17th ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and Van Rijsbergen, Cornelis J.}, publisher = {Springer Verlag, Heidelberg, Germany}, year = {1994}, address = {Dublin, Ireland}, pages = {282--289}, url = {\small http://www.acm.org/pubs/articles/proceedings/ir/188490/p282-hull/p282-hull.pdf}, abstract = {Latent Semantic Indexing (LSI) is a novel approach to information retrieval that attempts to model the underlying structure of term associations by transforming the traditional representation of documents as vectors of weighted term frequencies to a new coordinate space where both documents and terms are represented as linear combinations of underlying semantic factors. In previous research, LSI has produced a small improvement in retrieval performance. In this paper, we apply LSI to the routing task, which operates under the assumption that a sample of relevant and non-relevant documents is available to use in constructing the query. Once again, LSI slightly improves performance. However, when LSI is used is conduction with statistical classification, there is a dramatic improvement in performance.} } @inProceedings{Hull96, author = {David A. Hull and Jan O. Pedersen and Hinrich Sch{\"u}tze}, title = {Method combination for document filtering}, booktitle = {Proceedings of SIGIR-96, 19th ACM International Conference on Research and Development in Information Retrieval}, editor = {Hans-Peter Frei and Donna Harman and Peter Sch{\"{a}}uble and Ross Wilkinson}, publisher = {ACM Press, New York, US}, year = {1996}, address = {Z{\"{u}}rich, CH}, pages = {279--288}, url = {ftp://parcftp.xerox.com/pub/qca/papers/sigirfiltering96.ps}, abstract = {There is strong empirical and theoretic evidence that combination of retrieval methods can improve performance. In this paper, we systematically compare combination strategies in the context of document filtering, using queries from the Tipster reference corpus. We find that simple averaging strategies do indeed improve performance, but that direct averaging of probability estimates is not the correct approach. Instead, the probability estimates must be renormalized using logistic regression on the known relevance judgements. We examine more complex combination strategies but find them less successful due to the high correlations among our filtering methods which are optimized over the same training data and employ similar document representations.}, } @inProceedings{Hull98, author = {David A. Hull}, title = {The {TREC-7} filtering track: description and analysis}, booktitle = {Proceedings of TREC-7, 7th Text Retrieval Conference}, publisher = {National Institute of Standards and Technology, Gaithersburg, US}, editor = {Ellen M. Voorhees and Donna K. Harman}, year = {1998}, address = {Gaithersburg, US}, pages = {33--56}, url = {http://trec.nist.gov/pubs/trec7/papers/tr7filter/paper.ps}, abstract = {This article describes the experiments conducted in the TREC-7 filtering track, which consisted of three subtasks: adaptive filtering, batch filtering, and routing. The focus this year is on adaptive filtering, where the system begins with only the topic statement and must interactively adjust a filtering profile constructed from that topic in response to on-line feedback. In addition to motivating the task and describing the practical details of participating in the track, this document includes a detailed graphical presentation of the experimental results and provides a brief overall analysis of the performance data.}, } @inProceedings{Ipeirotis01, author = {Panagiotis G. Ipeirotis and Luis Gravano and Mehran Sahami}, title = {Probe, count, and classify: categorizing hidden Web databases}, booktitle = {Proceedings of SIGMOD-01, ACM International Conference on Management of Data}, editor = {Walid G. Aref}, publisher = {ACM Press, New York, US}, year = {2001}, address = {Santa Barbara, US}, pages = {67--78}, url = {http://doi.acm.org/10.1145/375663.375671}, abstract = {The contents of many valuable web-accessible databases are only accessible through search interfaces and are hence in-visible to traditional web ``crawlers''. Recent studies have estimated the size of this ''hidden web'' to be 500 billion pages, while the size of the ``crawlable'' web is only an es-timated two billion pages. Recently, commercial web sites have started to manually organize web-accessible databases into Yahoo!-like hierarchical classification schemes. In this paper, we introduce a method for automating this classi-fication process by using a small number of query probes. To classify a database, our algorithm does not retrieve or in-spect any documents or pages from the database, but rather just exploits the number of matches that each query probe generates at the database in question. We have conducted an extensive experimental evaluation of our technique over collections of real documents, including over one hundred web-accessible databases. Our experiments show that our system has low overhead and achieves high classification ac-curacy across a variety of databases.}, } @inProceedings{Ittner95, author = {David J. Ittner and Lewis, David D. and David D. Ahn}, title = {Text categorization of low quality images}, booktitle = {Proceedings of SDAIR-95, 4th Annual Symposium on Document Analysis and Information Retrieval}, publisher = {}, editor = {}, year = {1995}, address = {Las Vegas, US}, pages = {301--315}, url = {http://www.research.att.com/~lewis/papers/ittner95.ps}, abstract = {Categorization of text images into content-oriented classes would be a useful capability in a variety of document handling systems. Many methods can be used to categorize texts once their words are known, but OCR can garble a large proportion of words, particularly when low quality images are used. Despite this, we show for one data set that fax quality images can be categorized with nearly the same accuracy as the original text. Further, the categorization system can be trained on noisy OCR output, without need for the true text of any image, or for editing of OCR output. The use of a vector space classifier and training method robust to large feature sets, combined with discarding of low frequency OCR output strings are the key to our approach.}, } @inProceedings{Iwayama94, author = {Makoto Iwayama and Takenobu Tokunaga}, title = {A Probabilistic Model for Text Categorization: Based on a Single Random Variable with Multiple Values}, booktitle = {Proceedings of ANLP-94, 4th Conference on Applied Natural Language Processing}, publisher = {Association for Computational Linguistics, Morristown, US}, editor = {}, year = {1994}, address = {Stuttgart, DE}, pages = {162--167}, url = {}, abstract = {}, } @inProceedings{Iwayama95, author = {Makoto Iwayama and Takenobu Tokunaga}, title = {Cluster-based text categorization: a comparison of category search strategies}, booktitle = {Proceedings of SIGIR-95, 18th ACM International Conference on Research and Development in Information Retrieval}, editor = {Edward A. Fox and Peter Ingwersen and Raya Fidel}, publisher = {ACM Press, New York, US}, year = {1995}, address = {Seattle, US}, pages = {273--281}, url = {http://www.acm.org/pubs/articles/proceedings/ir/215206/p273-iwayama/p273-iwayama.pdf}, abstract = {Text categorization can be viewed as a process of category search, in which one or more categories for a test document are searched for by using given training documents with known categories. A cluster based search with a probabilistic clustering algorithm is proposed and evaluated on two data sets. The efficiency, effectiveness, and noise tolerance of this search strategy were confirmed to be better than those of a full search, a category based search, and a cluster based search with nonprobabilistic clustering.}, } @inProceedings{Iwayama95a, author = {Makoto Iwayama and Takenobu Tokunaga}, title = {Hierarchical Bayesian clustering for automatic text classification}, booktitle = {Proceedings of IJCAI-95, 14th International Joint Conference on Artificial Intelligence}, editor = {Chris E. Mellish}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, year = {1995}, address = {Montreal, CA}, pages = {1322--1327}, url = {}, abstract = {Text classification, the grouping of texts into several clusters, has been used as a means of improving both the efficiency and the effectiveness of text retrieval/categorization. In this paper we propose a hierarchical clustering algorithm that constructs a set of clusters having the maximum Bayesian posterior probability, the probability that the given texts are classified into clusters. We call the algorithm Hierarchical Bayesian Clustering (HBC). The advantages of HBC are experimentally verified from several viewpoints. HBC can reconstruct the original clusters more accurately than other non-probabilistic algorithms. When a probabilistic text categorization is extended to a cluster-based one, the use of HBC offers better performance than the use of non-probabilistic algorithms.}, } @inProceedings{Iwazume96, author = {Michiaki Iwazume and Hideaki Takeda and Toyoaki Nishida}, title = {Ontology-Based Information Gathering and Text Categorization from the Internet}, booktitle = {Proceedings of IEA/AIE-96, 9th International Conference in Industrial and Engineering Applications of Artificial Intelligence and Expert Systems}, editor = {}, publisher = {}, year = {1996}, address = {Fukuoka, JP}, pages = {305--314}, url = {}, abstract = {}, } @inProceedings{Iyer00, author = {Raj D. Iyer and David D. Lewis and Robert E. Schapire and Yoram Singer and Amit Singhal}, title = {Boosting for Document Routing}, booktitle = {Proceedings of CIKM-00, 9th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, address = {McLean, US}, editor = {Arvin Agah and Jamie Callan and Elke Rundensteiner}, year = {2000}, pages = {70--77}, url = {http://www.cs.huji.ac.il/~singer/papers/rankboost.ps.gz}, abstract = {RankBoost is a recently proposed algorithm for learning ranking functions. It is simple to implement and has strong justifications from computational learning theory. We describe the algorithm and present experimental results on applying it to the document routing problem. The first set of results applies RankBoost to a text representation produced using modern term weighting methods. Performance of RankBoost is somewhat inferior to that of a state-of-the-art routing algorithm which is, however, more complex and less theoretically justified than RankBoost. RankBoost achieves comparable performance to the state-of-the-art algorithm when combined with feature or example selection heuristics. Our second set of results examines the behavior of RankBoost when it has to learn not only a ranking function but also all aspects of term weighting from raw data. Performance is usually, though not always, less good here, but the term weighting functions implicit in the resulting ranking functions are intriguing, and the approach could easily be adapted to mixtures of textual and nontextual data.}, } @inProceedings{Jacobs92, author = {Paul S. Jacobs}, title = {Joining statistics with NLP for text categorization}, booktitle = {Proceedings of ANLP-92, 3rd Conference on Applied Natural Language Processing}, publisher = {Association for Computational Linguistics, Morristown, US}, editor = {Marcia Bates and Oliviero Stock}, year = {1992}, address = {Trento, IT}, pages = {178--185}, url = {}, abstract = {Automatic news categorization systems have produced high accuracy, consistency, and flexibility using some natural language processing techniques. These knowledge-based categorization methods are more powerful and accurate than statistical techniques. However, the phrasal pre-processing and pattern matching methods that seem to work for categorization have the disadvantage of requiring a fair amount of knowledge-encoding by human beings. In addition, they work much better at certain tasks, such as identifying major events in texts, than at others, such as determining what sort of business or product is involved in a news event. Statistical methods for categorization, on the other hand, are easy to implement and require little or no human customization. But they don't offer any of the benefits of natural language processing, such as the ability to identify relationships and enforce linguistic constraints. The authors' approach has been to use statistics in the knowledge acquisition component of a linguistic pattern-based categorization system, using statistical methods, for example, to associate words with industries and identify phrases that information about businesses or products. Instead of replacing knowledge-based methods with statistics, statistical training replaces knowledge engineering. This has resulted in high accuracy, shorter customization time, and good prospects for the application of the statistical methods to problems in lexical acquisition.}, } @article{Jacobs93, author = {Paul S. Jacobs}, title = {Using Statistical Methods to Improve Knowledge-Based News Categorization}, journal = {IEEE Expert}, year = {1993}, number = {2}, volume = {8}, pages = {13--23}, url = {}, abstract = {}, } @inProceedings{Jo99, author = {Taeho C. Jo}, title = {Text categorization with the concept of fuzzy set of informative keywords}, booktitle = {Proceedings of FUZZ-IEEE'99, IEEE International Conference on Fuzzy Systems}, editor = {}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, address = {Seoul, KR}, pages = {609--614}, year = {1999}, url = {}, abstract = {Text categorization is the procedure of assigning a category to a particular document among predefined categories. Informative keywords are the ones which reflect the contents of a document. A document includes informative keywords and non-informative keywords. Mainly non-informative keywords play the roles of grammatical functions in sentences; such keywords, what are called functional keywords, reflect its contents very little, so they should be removed in the process of document indexing. The discrimination between informative keywords and functional keywords is not crisp. In the process of document indexing, a document is represented as a set of informative keywords. In this paper, it is proposed that a document be represented into a fuzzy set of informative keywords, instead of a crisp set of informative keywords. The experiments of the categorization of news articles show that the proposed schemes of text categorization outperform the schemes with crisp sets.}, } @inCollection{Jo99a, author = {Taeho C. Jo}, title = {News article classification based on categorical points from keywords in backdata}, booktitle = {Computational Intelligence for Modelling, Control and Automation}, editor = {Masoud Mohammadian}, publisher = {IOS Press}, address = {Amsterdam, NL}, pages = {211--214}, year = {1999}, url = {}, abstract = {A scheme of automatic document classification is presented. Previously, documents have been classified according to their contents manually. Therefore, it is very costly to assign a category to them because a human investigates their contents. As the amount of data stored in storage media is increased exponentially, it becomes necessary to store documents according to their category, to access them easily. Automatic text classification is needed to store documents like that. Before performing text classification, back data should be constructed. The back data stores the information about keywords: the frequency for each category, the number of documents for each category. A document is represented with a list of keywords. Categorical points to each category are computed by summing the frequency of each keyword from back data, or the number of documents from it. The category that contains the largest categorical points is selected as the category of a document. In the results of an experiment with news article classification, precision is about 98\%.}, } @inCollection{Jo99b, author = {Taeho C. Jo}, title = {News articles classification based on representative keywords of categories}, booktitle = {Computational Intelligence for Modelling, Control and Automation}, editor = {Masoud Mohammadian}, publisher = {IOS Press}, address = {Amsterdam, NL}, pages = {194--198}, year = {1999}, url = {}, abstract = {A scheme of automatic document classification is presented. So far, documents have been classified according to their contents manually. Therefore, it is very costly to assign a category for them because humans investigate their contents. As the amount of data stored in storage media is increased exponentially, it becomes necessary to store documents according to their category, to access them easily. Automatic text classification is necessary to store documents like that. The scheme for automatic text classification proposed in the paper, is based on document indexing, where a document is represented as a list of keywords. The number of common keywords between keywords from the document itself and representative keywords from back data classifies documents. As an example, the proposed scheme is applied to the classification of news articles into 3 categories: politics, sports, and business. The measurements of performance evaluation are: classification rate, correctness rate, and classified correctness rate.}, } @inProceedings{Joachims00, author = {Thorsten Joachims}, title = {Estimating the Generalization Performance of a SVM Efficiently}, booktitle = {Proceedings of ICML-00, 17th International Conference on Machine Learning}, editor = {Pat Langley}, year = {2000}, address = {Stanford, US}, pages = {431--438}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www-ai.cs.uni-dortmund.de/DOKUMENTE/joachims_00a.pdf}, abstract = {This paper proposes and analyzes an efficient and effective approach for estimating the generalization performance of a support vector machine (SVM) for text classification. Without any computation-intensive resampling, the new estimators are computationally much more efficient than cross-validation or bootstrapping. They can be computed at essentially no extra cost immediately after training a single SVM. Moreover, the estimators developed here address the special performance measures needed for evaluating text classifiers. They can be used not only to estimate the error rate, but also to estimate recall, precision, and F1. A theoretical analysis and experiments show that the new method can effectively estimate the performance of SVM text classifiers in an efficient way.}, } @inProceedings{Joachims01b, author = {Thorsten Joachims and Nello Cristianini and John Shawe-Taylor}, title = {Composite Kernels for Hypertext Categorisation}, booktitle = {Proceedings of ICML-01, 18th International Conference on Machine Learning}, editor = {Carla Brodley and Andrea Danyluk}, address = {Williams College, US}, year = {2001}, pages = {250--257}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.cs.cornell.edu/People/tj/publications/joachims_etal_01a.pdf}, abstract = {Kernels are problem-specific functions that act as an interface between the learning system and the data. While it is well-known when the combination of two kernels is again a valid kernel, it is an open question if the resulting kernel will perform well. In particular, in which situations can a combination of kernel be expected to perform better than its components considered separately? Intuitively, one would like each of the two kernels to contribute information that is not available to the other. This characterization hence must consider the data at hand, both the kernels and also the task, that is the information given by the labels. We investigate this problem by looking at the task of designing kernels for hypertext classification, where both words and links information can be exploited. Firstly we introduce a novel kernel, whose Gram matrix is the well known co-citation matrix from bibliometrics, and demonstrate on real data that it has a good performance. Then we study the problem of combining it with a standard bag of words kernel. We provide sufficient conditions that indicate when an improvement can be expected, highlighting and formalising the notion of ``independent kernels''. Experimental results confirm the predictions of the theory in the hypertext domain.}, } @inProceedings{Joachims01c, author = {Thorsten Joachims}, title = {A Statistical Learning Model of Text Classification with Support Vector Machines}, booktitle = {Proceedings of SIGIR-01, 24th ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and David J. Harper and Donald H. Kraft and Justin Zobel}, publisher = {ACM Press, New York, US}, address = {New Orleans, US}, year = {2001}, pages = {128--136}, url = {http://www.cs.cornell.edu/People/tj/publications/joachims_01a.pdf}, abstract = {This paper develops a theoretical learning model of text classification for Support Vector Machines (SVMs). It connects the statistical properties of text-classification tasks with the generalization performance of a SVM in a quantitative way. Unlike conventional approaches to learning text classifiers, which rely primarily on empirical evidence, this model explains why and when SVMs perform well for text classification. In particular, it addresses the following questions: Why can support vector machines handle the large feature spaces in text classification effectively? How is this related to the statistical properties of text? What are sufficient conditions for applying SVMs to text-classification problems successfully?}, } @article{Joachims02, author = {Thorsten Joachims and Fabrizio Sebastiani}, title = {Guest editors' introduction to the special issue on automated text categorization}, journal = {Journal of Intelligent Information Systems}, year = {2002}, note = {Special Issue on Automated Text Categorization}, volume = {18}, number = {2/3}, pages = {103--105}, url = {http://www.wkap.nl/article.pdf?391241}, } @book{Joachims02a, author = {Thorsten Joachims}, title = {Learning to Classify Text using Support Vector Machines}, publisher = {Kluwer Academic Publishers}, address = {Dordrecht, NL}, year = {2002}, } @inProceedings{Joachims97, author = {Thorsten Joachims}, title = {A probabilistic analysis of the {Rocchio} algorithm with {TFIDF} for text categorization}, booktitle = {Proceedings of ICML-97, 14th International Conference on Machine Learning}, editor = {Douglas H. Fisher}, year = {1997}, address = {Nashville, US}, pages = {143--151}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www-ai.cs.uni-dortmund.de/DOKUMENTE/joachims_97a.ps.gz}, abstract = {The Rocchio relevance feedback algorithm is one of the most popular and widely applied learning methods from information retrieval. Here, a probabilistic analysis of this algorithm is presented in a text categorization framework. The analysis gives theoretical insight into the heuristics used in the Rocchio algorithm, particularly the word weighting scheme and the similarity metric. It also suggests improvements which lead to a probabilistic variant of the Rocchio classifier. The Rocchio classifier, its probabilistic variant, and a naive Bayes classifier are compared on six text categorization tasks. The results show that the probabilistic algorithms are preferable to the heuristic Rocchio classifier not only because they are more well-founded, but also because they achieve better performance.}, } @inProceedings{Joachims97b, author = {Thorsten Joachims and Dayne Freitag and Tom M. Mitchell}, title = {{\sc WebWatcher}: a tour guide for the Word Wide Web}, booktitle = {Proceedings of IJCAI-97, 15th International Joint Conference on Artificial Intelligence}, editor = {Martha E. Pollack}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, year = {1997}, address = {Nagoya, JP}, pages = {770--775}, url = {http://www.cs.cmu.edu/afs/cs/user/dayne/www/ps/ijcai97.ps.Z}, abstract = {We describe WebWatcher as a tour guide agent for the web, the learning algorithms used by WebWatcher, experimental results based on learning from thousands of users, and lessons learned from this case study of tour guide agents.}, } @inProceedings{Joachims98, author = {Thorsten Joachims}, title = {Text categorization with support vector machines: learning with many relevant features}, booktitle = {Proceedings of ECML-98, 10th European Conference on Machine Learning}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1398}, editor = {Claire N{\'{e}}dellec and C{\'{e}}line Rouveirol}, address = {Chemnitz, DE}, pages = {137--142}, year = {1998}, url = {http://www-ai.cs.uni-dortmund.de/DOKUMENTE/joachims_98a.ps.gz}, abstract = {The paper explores the use of Support Vector Machines (SVMs) for learning text classifiers from examples. It analyzes the particular properties of learning with text data and identifies why SVMs are appropriate for this task. Empirical results support the theoretical findings. SVMs achieve substantial improvements over the currently best performing methods and behave robustly over a variety of different learning tasks. Furthermore, they are fully automatic, eliminating the need for manual parameter tuning.}, } @inProceedings{Joachims99, author = {Thorsten Joachims}, title = {Transductive Inference for Text Classification using Support Vector Machines}, booktitle = {Proceedings of ICML-99, 16th International Conference on Machine Learning}, editor = {Ivan Bratko and Saso Dzeroski}, year = {1999}, address = {Bled, SL}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, pages = {200--209}, url = {http://www-ai.cs.uni-dortmund.de/DOKUMENTE/joachims_99c.ps.gz}, abstract = {This paper introduces transductive support vector machines (TSVMs) for text classification. While regular support vector machines (SVMs) try to induce a general decision function for a learning task, TSVMs take into account a particular test set and try to minimize misclassifications of just those particular examples. The paper presents an analysis of why TSVMs are well suited for text classification. These theoretical findings are supported by experiments on three test collections. The experiments show substantial improvements over inductive methods, especially for small training sets, cutting the number of labeled training examples down to a 20th on some tasks. This work also proposes an algorithm for training TSVMs efficiently, handling 10,000 examples and more.}, } @article{Juan02, author = {Juan, Alfons and Vidal, Enrique}, title = {On the use of Bernoulli mixture models for text classification}, journal = {Pattern Recognition}, year = {2002}, volume = {35}, number = {12}, pages = {2705--2710}, url = {}, abstract = {Mixture modelling of class-conditional densities is a standard pattern recognition technique. Although most research on mixture models has concentrated on mixtures for continuous data, emerging pattern recognition applications demand extending research efforts to other data types. This paper focuses on the application of mixtures of multivariate Bernoulli distributions to binary data. More concretely, a text classification task aimed at improving language modelling for machine translation is considered.}, } @inProceedings{Junker00, author = {Markus Junker and Michaell Sintek and Matthias Rinck}, title = {Learning for text categorization and information extraction with ILP}, booktitle = {Proceedings of the 1st Workshop on Learning Language in Logic}, editor = {Cussens, James and Saso Dzeroski}, year = {2000}, address = {Bled, SL}, pages = {247--258}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1925}, url = {}, abstract = {Text categorization (TC) and information extraction (IE) are two important goals of natural language processing. While hand-crafting rules for both tasks has a long tradition, learning approaches used to gain much interest in the past. Since in both tasks text as a sequence of words is of crucial importance, propositional learners have strong limitations, Although viewing learning for TC and IE as inductive logic programming (ILP) problems is obvious, most approaches rather use proprietary formalisms. In this paper, we provide a solid basis for the application of ILP methods to these learning problems. We introduce three basic types (namely a type for text, one for words and one for positions in texts) and three simple predicate definitions over these types which enable us to write TC and IE rules as logic programs. Based on the proposed representation, we present an approach to the problem of learning rules for TC and IE in terms of ILP. We conclude by comparing our approach of representing texts and rules as logic programs to others.}, } @inProceedings{Junker01, author = {Markus Junker and Andreas Dengel}, title = {Preventing Overfitting in Learning Text Patterns for Document Categorization}, booktitle = {Proceedings of ICAPR-01, 2nd International Conference on Advances in Pattern Recognition}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2013}, editor = {Sameer Singh and Nabeel A. Murshed and Walter Kropatsch}, address = {Rio De Janeiro, BR}, year = {2001}, pages = {137--146}, url = {http://link.springer.de/link/service/series/0558/papers/2013/20130137.pdf}, abstract = {There is an increasing interest in categorizing texts using learning algorithms. While the majority of approaches rely on learning linear classifiers, there is also some interest in describing document categories by text patterns. We introduce a model for learning patterns for text categorization (the LPT-model) that does not rely on an attribute-value representation of documents but represents documents essentially "as they are". Based on the LPT-model, we focus on learning patterns within a relatively simple pattern language. We compare different search heuristics and pruning methods known from various symbolic rule learners on a set of representative text categorization problems. The best results were obtained using the m-estimate as search heuristics combined with the likelihood-ratio-statics for pruning. Even better results can be obtained, when replacing the likelihood-ratio-statics by a new measure for pruning; this we call l-measure. In contrast to conventional measures for pruning, the l-measure takes into account properties of the search space.}, } @inProceedings{Junker97, author = {Markus Junker and Andreas Abecker}, title = {Exploiting Thesaurus Knowledge in Rule Induction for Text Classification}, booktitle = {Proceedings of RANLP-97, 2nd International Conference on Recent Advances in Natural Language Processing}, publisher = {}, editor = {Ruslan Milkov and Nicolas Nicolov and Nilokai Nikolov}, address = {Tzigov Chark, BL}, pages = {202--207}, year = {1997}, url = {http://www.dfki.uni-kl.de/~junker/download/ranlp97.ps}, abstract = {Systems for learning text classifiers recently gained considerable interest. One technique to implement such systems is rule induction. While most other approaches rely on a relatively simple document representation and do not make use of any background knowledge, rule induction algorithms offer a good potential for improvements in both of these areas. In this paper, we show how an operator-based view of rule induction enables the easy integration of a thesaurus as background knowledge. Results with an algorithm extended by thesaurus knowledge are presented and interpreted. The interpretation shows the strengths and weaknesses of using thesaurus knowledge and gives hints for future research.}, } @article{Junker98, author = {Markus Junker and Rainer Hoch}, title = {An experimental evaluation of OCR text representations for learning document classifiers}, journal = {International Journal on Document Analysis and Recognition}, pages = {116--122}, year = {1998}, number = {2}, volume = {1}, url = {http://link.springer.de/link/service/journals/10032/papers/8001002/80010116.ps.gz}, abstract = {In the literature, many feature types are proposed for document classification. However, an extensive and systematic evaluation of the various approaches has not yet been done. In particular, evaluations on OCR documents are very rare. In this paper we investigate seven text representations based on n-grams and single words. We compare their effectiveness in classifying OCR texts and the corresponding correct ASCII texts in two domains: business letters and abstracts of technical reports. Our results indicate that the use of n-grams is an attractive technique which can even compare to techniques relying on a morphological analysis. This holds for OCR texts as well as for correct ASCII texts.}, } @article{Kaban02, author = {Ata Kaban and Mark Girolami}, title = {A Dynamic Probabilistic Model to Visualise Topic Evolution in Text Streams}, journal = {Journal of Intelligent Information Systems}, year = {2002}, note = {Special Issue on Automated Text Categorization}, volume = {18}, number = {2/3}, pages = {107--125}, url = {http://www.wkap.nl/article.pdf?391242}, abstract = {We propose a novel probabilistic method, based on latent variable models, for unsupervised topographic visualisation of dynamically evolving, coherent textual information. This can be seen as a complementary tool for topic detection and tracking applications. This is achieved by the exploitation of the a priori domain knowledge available, that there are relatively homogeneous temporal segments in the data stream. In a different manner from topographical techniques previously utilized for static text collections, the topography is an outcome of the coherence in time of the data stream in the proposed model. Simulation results on both toy-data settings and an actual application on Internet chat line discussion analysis is presented by way of demonstration.}, } @inProceedings{Kao03, author = {Anne Kao and Lesley Quach and Steve Poteet and Steve Woods}, title = {User assisted text classification and knowledge management}, booktitle = {Proceedings of CIKM-03, 12th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {2003}, address = {New Orleans, US}, pages = {524--527}, url = {http://doi.acm.org/10.1145/956863.956964}, abstract = {While there are many aspects to managing corporate knowledge, one key issue is how to organize corporate documents into categories of interest. In this paper, we focus on using user assisted text classification in conjunction with a web portal, multiple document management systems and an ontology, to provide a powerful solution for organizing information about a company's technology. We propose a system that interacts with an author using an automatic text classifier to suggest controlled keywords to be used as metadata. The proposed approach does not require professional librarians or that the end users have extensive training. The use of a controlled vocabulary allows for a more consistent description of corporate documents, and promotes easier access by people across the company. It is easier to find similar documents which use different nomenclature. Finally, the interactive nature of the system results in a more correct and precise description of each document than a fully automatic system would.}, } @article{Kar78, author = {Gautam Kar and Lee J. White}, title = {A distance measure for automated document classification by sequential analysis}, journal = {Information Processing and Management}, pages = {57--69}, year = {1978}, number = {2}, volume = {14}, url = {}, abstract = {}, } @inProceedings{Karypis00, author = {George Karypis and Eui-Hong Han}, title = {Fast Supervised Dimensionality Reduction Algorithm with Applications to Document Categorization and Retrieval}, booktitle = {Proceedings of CIKM-00, 9th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, address = {McLean, US}, editor = {Arvin Agah and Jamie Callan and Elke Rundensteiner}, year = {2000}, pages = {12--19}, url = {ftp://ftp.cs.umn.edu/dept/users/kumar/cikm-ci.ps}, abstract = {Retrieval techniques based on dimensionality reduction, such as Latent Semantic Indexing (LSI), have been shown to improve the quality of the information being retrieved by capturing the latent meaning of the words present in the documents. Unfortunately, the high computational and memory requirements of LSI and its inability to compute an effective dimensionality reduction in a supervised setting limits its applicability. In this paper we present a fast supervised dimensionality reduction algorithm that is derived from the recently developed cluster-based unsupervised dimensionality reduction algorithms. We experimentally evaluate the quality of the lower dimensional spaces both in the context of document categorization and improvements in retrieval performance on a variety of different document collections. Our experiments show that the lower dimensional spaces computed by our algorithm consistently improve the performance of traditional algorithms such as C4.5, k-nearest-neighbor, and Support Vector Machines (SVM), by an average of 2\% to 7\%. Furthermore, the supervised lower dimensional space greatly improves the retrieval performance when compared to LSI.}, } @inProceedings{Kawatani02, author = {Takahiko Kawatani}, title = {Topic Difference Factor Extraction between Two Document Sets and its Application to Text Categorization}, booktitle = {Proceedings of SIGIR-02, 25th ACM International Conference on Research and Development in Information Retrieval}, editor = {Micheline Beaulieu and Ricardo Baeza-Yates and Sung Hyon Myaeng and Kalervo J{\"{a}}rvelin}, publisher = {ACM Press, New York, US}, address = {Tampere, FI}, year = {2002}, pages = {137--144}, url = {http://doi.acm.org/10.1145/564376.564402}, abstract = {To improve performance in text categorization, it is important to extract distinctive features for each class. This paper proposes topic difference factor analysis (TDFA) as a method to extract projection axes that reflect topic differences between two document sets. Suppose all sentence vectors that compose each document are projected onto projection axes. TDFA obtains the axes that maximize the ratio between the document sets as to the sum of squared projections by solving a generalized eigenvalue problem. The axes are called topic difference factors (TDF's). By applying TDFA to the document set that belongs to a given class and a set of documents that is misclassified as belonging to that class by an existent classifier, we can obtain features that take large values in the given class but small ones in other classes, as well as features that take large values in other classes but small ones in the given class. A classifier was constructed applying the above features to complement the kNN classifier. As the results, the micro averaged F1 measure for Reuters-21578 improved from 83.69 to 87.27\%.}, } @article{Kehagias03, title = {A Comparison of Word- and Sense-based Text Categorization Using Several Classification Algorithms}, author = {Athanasios Kehagias and Vassilios Petridis and Vassilis G. Kaburlasos and Pavlina Fragkou}, journal = {Journal of Intelligent Information Systems}, year = {2003}, volume = {21}, number = {3}, pages = {227--247}, url = {http://www.wkap.nl/article.pdf?391243}, abstract = {Most of the text categorization algorithms in the literature represent documents as collections of words. An alternative which has not been sufficiently explored is the use of word meanings, also known as senses. In this paper, using several algorithms, we compare the categorization accuracy of classifiers based on words to that of classifiers based on senses. The document collection on which this comparison takes place is a subset of the annotated Brown Corpus semantic concordance. A series of experiments indicates that the use of senses does not result in any significant categorization improvement.}, } @inProceedings{Kessler97, author = {Brett Kessler and Geoff Nunberg and Hinrich Sch{\"{u}}tze}, title = {Automatic detection of text genre}, booktitle = {Proceedings of ACL-97, 35th Annual Meeting of the Association for Computational Linguistics}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, editor = {Philip R. Cohen and Wolfgang Wahlster}, year = {1997}, address = {Madrid, ES}, pages = {32--38}, url = {ftp://parcftp.xerox.com/pub/qca/genre/paper.acl97.ps.Z}, abstract = {As the text databases available to users become larger and more heterogeneous, genre becomes increasingly important for computational linguistics as a complement to topical and structural principles of classification. We propose a theory of genres as bundles of facets, which correlate with various surface cues, and argue that genre detection based on surface cues is as successful as detection based on deeper structural properties.}, } @inProceedings{Khmelev03, author = {Dmitry V. Khmelev and William J. Teahan}, title = {A repetition based measure for verification of text collections and for text categorization}, booktitle = {Proceedings of SIGIR-03, 26th ACM International Conference on Research and Development in Information Retrieval}, editor = {Jamie Callan and Gordon Cormack and Charles Clarke and David Hawking and Alan Smeaton}, publisher = {ACM Press, New York, US}, address = {Toronto, CA}, year = {2003}, pages = {104--110}, url = {http://doi.acm.org/10.1145/860435.860456}, abstract = {We suggest a way for locating duplicates and plagiarisms in a text collection using an R-measure, which is the normalized sum of the lengths of all suffixes of the text repeated in other documents of the collection. The R-measure can be effectively computed using the suffix array data structure. Additionally, the computation procedure can be improved to locate the sets of duplicate or plagiarised documents. We applied the technique to several standard text collections and found that they contained a significant number of duplicate and plagiarised documents. Another reformulation of the method leads to an algorithm that can be applied to supervised multi-class categorization. We illustrate the approach using the recently available Reuters Corpus Volume 1 (RCV1). The results show that the method outperforms SVM at multi-class categorization, and interestingly, that results correlate strongly with compression-based methods.}, } @inProceedings{Kim00, author = {Yu-Hwan Kim and Shang-Yoon Hahn and Byoung-Tak Zhang}, title = {Text filtering by boosting naive Bayes classifiers}, booktitle = {Proceedings of SIGIR-00, 23rd ACM International Conference on Research and Development in Information Retrieval}, editor = {Nicholas J. Belkin and Peter Ingwersen and Mun-Kew Leong}, publisher = {ACM Press, New York, US}, address = {Athens, GR}, year = {2000}, pages = {168--175}, url = {http://www.acm.org/pubs/articles/proceedings/ir/345508/p168-kim/p168-kim.pdf}, abstract = {Several machine learning algorithms have recently been used for text categorization and filtering. In particular, boosting methods such as AdaBoost have shown good performance applied to real text data. However, most of existing boosting algorithms are based on classifiers that use binary-valued features. Thus, they do not fully make use of the weight information provided by standard term weighting methods. In this paper, we present a boosting-based learning method for text filtering that uses naive Bayes classifiers as a weak learner. The use of naive Bayes allows the boosting algorithm to utilize term frequency information while maintaining probabilistically accurate confidence ratio. Applied to TREC-7 and TREC-8 filtering track documents, the proposed method obtained a significant improvement in LF1, LF2, Fl and F3 measures compared to the best results submitted by other TREC entries.}, } @inProceedings{Kim04, author = {Sang-Bum Kim and Hae-Chang Rim}, title = {Recomputation of Class Relevance Scores for Improving Text Classification}, booktitle = {Proceedings of CICLING-04, 5th International Conference on Computational Linguistics and Intelligent Text Processing}, year = {2004}, editor = {Alexander F. Gelbukh}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Seoul, KO}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2945}, pages = {580--583}, url = {}, abstract = {}, } @article{Kim04a, author = {Kim, J. and Kim, M.}, title = {An evaluation of passage-based text categorization}, journal = {Journal of Intelligent Information Systems}, year = {2004}, volume = {23}, number = {1}, pages = {47--65}, url = {http://dx.doi.org/10.1023/B:JIIS.0000029670.53363.d0}, abstract = {Researches in text categorization have been confined to whole-document-level classification, probably due to lack of full-text test collections. However, full-length documents available today in large quantities pose renewed interests in text classification. A document is usually written in an organized structure to present its main topic(s). This structure can be expressed as a sequence of subtopic text blocks, or passages. In order to reflect the subtopic structure of a document, we propose a new passage-level or passage-based text categorization model, which segments a test document into several passages, assigns categories to each passage, and merges the passage categories to the document categories. Compared with traditional document-level categorization, two additional steps, passage splitting and category merging, are required in this model. Using four subsets of the Reuters text categorization test collection and a full-text test collection of which documents are varying from tens of kilobytes to hundreds, we evaluate the proposed model, especially the effectiveness of various passage types and the importance of passage location in category merging. Our results show simple windows are best for all test collections tested in these experiments. We also found that passages have different degrees of contribution to the main topic(s), depending on their location in the test document.}, } @inProceedings{Kindermann01, author = {J{\"{o}}rg Kindermann and Gerhard Paa{{\ss}} and Edda Leopold}, title = {Error Correcting Codes with Optimized Kullback-Leibler Distances for Text Categorization}, booktitle = {Proceedings of ECML-01, 12th European Conference on Machine Learning}, editor = {Luc De Raedt and Arno Siebes}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Freiburg, DE}, year = {2001}, pages = {266--275}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2168}, url = {http://link.springer.de/link/service/series/0558/papers/2168/21680266.pdf}, abstract = {We extend a multi-class categorization scheme proposed by Dietterich and Bakiri 1995 for binary classifiers, using error correcting codes. The extension comprises the computation of the codes by a simulated annealing algorithm and optimization of Kullback-Leibler (KL) category distances within the code-words. For the first time, we apply the scheme to text categorization with support vector machines (SVMs) on several large text corpora with more than 100 categories. The results are compared to 1-of-N coding (i.e.\ one SVM for each text category). We also investigate codes with optimized KL distance between the text categories which are merged in the code-words. We find that error correcting codes perform better than 1-of-N coding with increasing code length. For very long codes, the performance is in some cases further improved by KL-distance optimization.}, } @inProceedings{Klas00, author = {Klas, Claus-Peter and Fuhr, Norbert}, title = {A new Effective Approach for Categorizing Web Documents}, booktitle = {Proceedings of BCSIRSG-00, the 22nd Annual Colloquium of the British Computer Society Information Retrieval Specialist Group}, editor = {}, address = {Cambridge, UK}, year = {2000}, pages = {}, publisher = {}, url = {http://ls6-www.informatik.uni-dortmund.de/bib/fulltext/ir/Klas_Fuhr:00.ps.gz}, abstract = {Categorization of Web documents poses a new challenge for automatic classification methods. In this paper, we present the megadocument approach for categorization. For each category, all corresponding document texts from the training sample are concatenated to a megadocument, which is indexed using standard methods. In order to classify a new document, the most similar megadocument determines the category to be assigned. Our evaluations show that for Web collections, the megadocument method clearly outperformes other classification methods. In contrast, for the Reuters collection, we only achieve mediocre results. Thus, our method seems to be well suited for heterogeneous document collections.}, } @article{Klingbiel73, author = {Paul H. Klingbiel}, title = {Machine-aided indexing of technical literature}, journal = {Information Storage and Retrieval}, year = {1973}, volume = {9}, number = {2}, pages = {79--84}, url = {}, abstract = {To index successfully in the Defense Documentation Center's environment, an automated system must chose single words or phrases (dependent upon context) rapidly and economically. The automation of DDC's indexing has been machine-aided from its inception. A machine-aided indexing (MAI) system is described that indexes one million words of text per hour of CPU time. Grammatical errors do not exceed five per cent of the output, so human screening is satisfactorily low. The system could potentially scale up to an operational size of 10 million words of text per year - the equivalent of a dozen bibles or a third of the Encyclopedia Britannica. In a batch mode, the programs to accomplish this indexing would require no more than fifteen minutes of CPU time per week.}, } @article{Klingbiel73a, author = {Paul H. Klingbiel}, title = {A technique for machine-aided indexing}, journal = {Information Storage and Retrieval}, year = {1973}, volume = {9}, number = {9}, pages = {477--494}, url = {}, abstract = {Subject indexing of text can, in principle, be accomplished in many ways. The technique for machine-aided indexing (MAI) developed at the Defense Documentation Center (DDC) is illustrated on a randomly chosen abstract. Additional text is provided in coded form so that the reader can more fully explore this technique and form his own opinion of the applicability and versatility of this particular procedure. The DDC method for subject indexing is very close to operational status for a data base which grows at the rate of two million words of text per year.}, } @inProceedings{Klinkenberg00, author = {Ralf Klinkenberg and Thorsten Joachims}, title = {Detecting concept drift with support vector machines}, booktitle = {Proceedings of ICML-00, 17th International Conference on Machine Learning}, editor = {Pat Langley}, year = {2000}, address = {Stanford, US}, pages = {487--494}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www-ai.cs.uni-dortmund.de/DOKUMENTE/klinkenberg_joachims_2000a.pdf.gz}, abstract = {For many learning tasks where data is collected over an extended period of time, its underlying distribution is likely to change. A typical example is information filtering, i.e. the adaptive classification of documents with respect to a particular user interest. Both the interest of the user and the document content change over time. A filtering system should be able to adapt to such concept changes. This paper proposes a new method to recognize and handle concept changes with support vector machines. The method maintains a window on the training data. The key idea is to automatically adjust the window size so that the estimated generalization error is minimized. The new approach is both theoretically well-founded as well as effective and efficient in practice. Since it does not require complicated parameterization, it is simpler to use and more robust than comparable heuristics. Experiments with simulated concept drift scenarios based on real-world text data compare the new method with other window management approaches. We show that it can effectively select an appropriate window size in a robust way.}, } @inProceedings{Knorz82, author = {Knorz, Gerhard}, title = {A decision theory approach to optimal automated indexing}, booktitle = {Proceedings of SIGIR-82, 5th ACM International Conference on Research and Development in Information Retrieval}, year = {1982}, editor = {Gerard Salton and Hans-Jochen Schneider}, pages = {174--193}, address = {Berlin, DE}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 146}, url = {}, abstract = {}, } @inProceedings{Ko00, author = {Youngjoong Ko and Jungyun Seo}, title = {Automatic Text Categorization by Unsupervised Learning}, booktitle = {Proceedings of COLING-00, the 18th International Conference on Computational Linguistics}, year = {2000}, editor = {}, pages = {}, address = {Saarbr{\"{u}}cken, DE}, url = {http://nlp3.korea.ac.kr/proceeding/coling2000/COLING/ps/066.ps}, abstract = {The goal of text categorization is to classify documents into a certain number of pre-defined categories. The previous works in this area have used a large number of labeled training documents for supervised learning. One problem is that it is difficult to create the labeled training documents. While it is easy to collect the unlabeled documents, it is not so easy to manually categorize them for creating training documents. In this paper, we propose an unsupervised learning method to overcome these difficulties. The proposed method divides the documents into sentences, and categorizes each sentence using keyword lists of each category and sentence similarity measure. And then, it uses the categorized sentences for training. The proposed method shows a similar degree of performance, compared with the traditional supervised learning methods. Therefore, this method can be used in areas where low-cost text categorization is needed. It also can be used for creating training documents.}, } @inProceedings{Ko02, author = {Youngjoong Ko and Jinwoo Park and Jungyun Seo}, title = {Automatic Text Categorization using the Importance of Sentences}, booktitle = {Proceedings of COLING-02, the 19th International Conference on Computational Linguistics}, year = {2002}, editor = {}, pages = {}, address = {Taipei, TW}, url = {http://acl.ldc.upenn.edu/coling2002/proceedings/data/area-28/co-201.pdf}, abstract = {This paper proposes a new approach for text categorization, based on a feature projection technique. In our approach, training data are represented as the projections of training documents on each feature. The voting for a classification is processed on the basis of individual feature projections. The final classification of test documents is determined by a majority voting from the individual classifications of each feature. Our empirical results show that the proposed approach, Text Categorization using Feature Projections (TCFP), outperforms k-NN, Rocchio, and Naive Bayes. Most of all, TCFP is about one hundred times faster than k-NN. Since TCFP algorithm is very simple, its implementation and training process can be done very easily. For these reasons, TCFP can be a useful classifier in the areas, which need a fast and high-performance text categorization task.}, } @inProceedings{Ko02a, author = {Youngjoong Ko and Jungyun Seo}, title = {Text Categorization using Feature Projections}, booktitle = {Proceedings of COLING-02, the 19th International Conference on Computational Linguistics}, year = {2002}, editor = {}, pages = {}, address = {Taipei, TW}, url = {http://acl.ldc.upenn.edu/coling2002/proceedings/data/area-28/co-269.pdf}, abstract = {Automatic text categorization is a problem of automatically assigning text documents to predefined categories. In order to classify text documents, we must extract good features from them. In previous research, a text document is commonly represented by the term frequency and the inverted document frequency of each feature. Since there is a difference between important sentences and unimportant sentences in a document, the features from more important sentences should be considered more than other features. In this paper, we measure the importance of sentences using text summarization techniques. Then a document is represented as a vector of features with different weights according to the importance of each sentence. To verify our new method, we conducted experiments on two language newsgroup data sets: one written by English and the other written by Korean. Four kinds of classifiers were used in our experiments: Naive Bayes, Rocchio, k-NN, and SVM. We observed that our new method made a significant improvement in all classifiers and both data sets.}, } @article{Ko04, author = {Youngjoong Ko and Jinwoo Park and Jungyun Seo}, title = {Improving Text Categorization using the Importance of Sentences}, journal = {Information Processing and Management}, year = {2004}, volume = {40}, number = {1}, pages = {65--79}, url = {}, abstract = {}, } @article{Ko04a, author = {Youngjoong Ko and Jinwoo Park and Jungyun Seo}, title = {Using the feature projection technique based on a normalized voting method for text classification}, journal = {Information Processing and Management}, year = {2004}, volume = {40}, number = {2}, pages = {191--208}, url = {}, abstract = {}, } @inProceedings{Koehn02, author = {Philipp Koehn}, title = {Combining Multiclass Maximum Entropy Text Classifiers with Neural Network Voting}, booktitle = {Proceedings of PorTAL-02, 3rd International Conference on Advances in Natural Language Processing}, year = {2002}, editor = {Elisabete Ranchod and Nuno J. Mamede}, pages = {125--132}, address = {Faro, PT}, url = {http://link.springer.de/link/service/series/0558/papers/2389/23890125.pdf}, abstract = {We improve a high-accuracy maximum entropy classifier by combining an ensemble of classifiers with neural network voting. In our experiments we demonstrate significantly superior performance both over a single classifier as well as over the use of the traditional weighted-sum voting approach. Specifically, we apply this to a maximum entropy classifier on a large scale multi-class text categorization task: the online job directory Flipdog with over half a million jobs in 65 categories.}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2389}, } @inProceedings{Kolcz01, author = {Aleksander Kolcz and Vidya Prabakarmurthi and Jugal K. Kalita}, title = {String Match and Text Extraction: Summarization as feature selection for text categorization}, booktitle = {Proceedings of CIKM-01, 10th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {Henrique Paques and Ling Liu and David Grossman}, year = {2001}, address = {Atlanta, US}, pages = {365--370}, url = {http://doi.acm.org/10.1145/502585.502647}, abstract = {We address the problem of evaluating the effectiveness of summarization techniques for the task of document categorization. It is argued that for a large class of automatic categorization algorithms, extraction-based document categorization can be viewed as a particular form of feature selection performed on the full text of the document and, in this context, its impact can be compared with state-of-the-art feature selection techniques especially devised to provide good categorization performance. Such a framework provides for a better assessment of the expected performance of a categorizer if the compression rate of the summarizer is known.}, } @inProceedings{Koller97, author = {Daphne Koller and Mehran Sahami}, title = {Hierarchically classifying documents using very few words}, booktitle = {Proceedings of ICML-97, 14th International Conference on Machine Learning}, editor = {Douglas H. Fisher}, year = {1997}, address = {Nashville, US}, pages = {170--178}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://robotics.stanford.edu/users/sahami/papers-dir/ml97-hier.ps}, abstract = {The proliferation of topic hierarchies for text documents has resulted in a need for tools that automatically classify new documents within such hierarchies. Existing classification schemes which ignore the hierarchical structure and treat the topics as separate classes are often inadequate in text classification where the there is a large number of classes and a huge number of relevant features needed to distinguish between them. We propose an approach that utilizes the hierarchical topic structure to decompose the classification task into a set of simpler problems, one at each node in the classification tree. As we show, each of these smaller problems can be solved accurately by focusing only on a very small set of features, those relevant to the task at hand. This set of relevant features varies widely throughout the hierarchy, so that, while the overall relevant feature set may be large, each classifier only examines a small subset. The use of reduced feature sets allows us to utilize more complex (probabilistic) models, without encountering many of the standard computational and robustness difficulties.}, } @inProceedings{Kongovi02, author = {Madhusudhan Kongovi and Juan Carlos Guzman and Venu Dasigi}, title = {Text Categorization: An experiment using Phrases}, booktitle = {Proceedings of ECIR-02, 24th European Colloquium on Information Retrieval Research}, editor = {Fabio Crestani and Mark Girolami and Van Rijsbergen, Cornelis J.}, year = {2002}, address = {Glasgow, UK}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2291}, pages = {213--228}, url = {http://link.springer.de/link/service/series/0558/papers/2291/22910213.pdf}, abstract = {Typical text classifiers learn from example and training documents that have been manually categorized. In this research, our experiment dealt with the classification of news wire articles using category profiles. We built these profiles by selecting feature words and phrases from the training documents. For our experiments we decided on using the text corpus Reuters-21578. We used precision and recall to measure the effectiveness of our classifier. Though our experiments with words yielded good results, we found instances where the phrase-based approach produced more effectiveness. This could be due to the fact that when a word along with its adjoining word - a phrase - is considered towards building a category profile, it could be a good discriminator. This tight packaging of word pairs could bring in some semantic value. The packing of word pairs also filters out words occurring frequently in isolation that do not bear much weight towards characterizing that category.}, } @article{Koppel02, author = {Koppel, Moshe and Argamon, Shlomo and Shimoni, Anat R.}, title = {Automatically categorizing written texts by author gender}, journal = {Literary and Linguistic Computing}, year = {2002}, number = {4}, volume = {17}, pages = {401--412}, url = {http://www3.oup.co.uk/litlin/hdb/Volume_17/Issue_04/pdf/170401.pdf}, abstract = {The problem of automatically determining the gender of a document's author would appear to be a more subtle problem than those of categorization by topic or authorship attribution. Nevertheless, it is shown that automated text categorization techniques can exploit combinations of simple lexical and syntactic features to infer the gender of the author of an unseen formal written document with approximately 80 per cent accuracy. The same techniques can be used to determine if a document is fiction or non-fiction with approximately 98 per cent accuracy.}, } @inProceedings{Kosmynin96, author = {Arkadi Kosmynin and Ian Davidson}, title = {Using background contextual knowledge for documents representation}, booktitle = {Proceedings of PODP-96, 3rd International Workshop on Principles of Document Processing}, editor = {Charles K. Nicholas and Derick Wood}, year = {1996}, address = {Palo Alto, CA}, pages = {123--133}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1293}, url = {}, abstract = {We describe our approach to document representation that captures contextual dependencies between terms in a corpus and makes use of these dependencies to represent documents. We have tried our representation scheme for automatic document categorisation on the Reuters' test set of documents. We achieve a precision recall break even point of 84\% which is comparable to the best known published results. Our approach acts as a feature selection technique that is an alternative to applying the techniques from machine learning and numerical taxonomy.}, } @inProceedings{Koster03, author = {Cornelis H. Koster and Mark Seutter}, title = {Taming wild phrases}, booktitle = {Proceedings of ECIR-03, 25th European Conference on Information Retrieval}, publisher = {Springer Verlag, Heidelberg, DE}, editor = {Fabrizio Sebastiani}, address = {Pisa, IT}, year = {2003}, pages = {161--176}, url = {http://www.cs.kun.nl/~kees/peking/ecir03.pdf}, abstract = {In this paper the suitability of different document representations for automatic document classification is compared, investigating a whole range of representations between bag-of-words and bag-of-phrases. We look at some of their statistical properties, and determine for each representation the optimal choice of classification parameters and the effect of Term Selection. Phrases are represented by an abstraction called Head/Modifier pairs. Rather than just throwing phrases and keywords together, we shall start with pure HM pairs and gradually add more keywords to the document representation. We use the classification on keywords as the baseline, which we compare with the contribution of the pure HM pairs to classification accuracy, and the incremental contributions from heads and modifiers. Finally, we measure the accuracy achieved with all words and all HM pairs combined, which turns out to be only marginally above the baseline. We conclude that even the most careful term selection cannot overcome the differences in Document Frequency between phrases and words, and propose the use of term clustering to make phrases more cooperative.}, } @article{Krier02, author = {Marc Krier and Francesco Zacc{\`a}}, title = {Automatic categorization applications at the European Patent Office}, journal = {World Patent Information}, year = {2002}, volume = {24}, number = {}, pages = {187--196}, url = {}, abstract = {}, } @inProceedings{Krishnapuram03, author = {Raghu Krishnapuram and Krishna Chitrapura and Sachindra Joshi}, title = {Classification of Text Documents Based on Minimum System Entropy}, booktitle = {Proceedings of ICML-03, 20th International Conference on Machine Learning}, editor = {}, year = {2003}, address = {Washington, DC}, pages = {}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {}, abstract = {}, } @inProceedings{Kumaran04, author = {Giridhar Kumaran and James Allan}, title = {Text classification and named entities for new event detection}, booktitle = {Proceedings of SIGIR-04, 27th ACM International Conference on Research and Development in Information Retrieval}, editor = {Kalervo J{\"{a}}rvelin and James Allan and Peter Bruza and Mark Sanderson}, publisher = {ACM Press, New York, US}, address = {Sheffield, UK}, year = {2004}, pages = {297--304}, url = {http://doi.acm.org/10.1145/1008992.1009044}, abstract = {New Event Detection is a challenging task that still offers scope for great improvement after years of effort. In this paper we show how performance on New Event Detection (NED) can be improved by the use of text classification techniques as well as by using named entities in a new way. We explore modifications to the document representation in a vector space-based NED system. We also show that addressing named entities preferentially is useful only in certain situations. A combination of all the above results in a multi-stage NED system that performs much better than baseline single-stage NED systems.}, } @inProceedings{Kwok98, author = {James T. Kwok}, title = {Automated text categorization using support vector machine}, booktitle = {Proceedings of ICONIP'98, 5th International Conference on Neural Information Processing}, editor = {}, year = {1998}, address = {Kitakyushu, JP}, pages = {347--351}, url = {http://www.comp.hkbu.edu.hk/7Ejamesk/papers/iconip98.ps.gz}, abstract = {In this paper, we study the use of support vector machine in text categorization. Unlike other machine learning techniques, it allows easy incorporation of new documents into an existing trained system. Moreover, dimension reduction, which is usually imperative, now becomes optional. Thus, SVM adapts efficiently in dynamic environments that require frequent additions to the document collection. Empirical results on the Reuters-22173 collection are also discussed.}, } @article{Kwon03, author = {Oh-Woog Kwon and Jong-Hyeok Lee}, title = {Text categorization based on {k}-nearest neighbor approach for Web site classification}, journal = {Information Processing and Management}, year = {2003}, volume = {39}, number = {1}, pages = {25--44}, url = {}, abstract = {}, } @inProceedings{Kwon99, author = {Oh-Woog Kwon and Sung-Hwa Jung and Jong-Hyeok Lee and Geunbae Lee}, title = {Evaluation of Category Features and Text Structural Information on a Text Categorization Using Memory Based Reasoning}, booktitle = {Proceedings of ICCPOL-99, 18th International Conference on Computer Processing of Oriental Languages}, editor = {}, year = {1999}, address = {Tokushima, JP}, pages = {153--158}, url = {}, abstract = {}, } @inProceedings{Labrou99, author = {Yannis Labrou and Tim Finin}, title = {{{\sc Yahoo!}} as an ontology: using {{\sc Yahoo!}}\ categories to describe documents}, booktitle = {Proceedings of CIKM-99, 8th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {1999}, address = {Kansas City, US}, pages = {180--187}, url = {http://www.acm.org/pubs/articles/proceedings/cikm/319950/p180-labrou/p180-labrou.pdf}, abstract = {We suggest that one (or a collection) of names of {{\sc Yahoo!}}\ (or any other WWW indexer's) categories can be used to describe the content of a document. Such categories offer a standardized and universal way for referring to or describing the nature of real world objects, activities, documents and so on, and may be used (we suggest) to semantically characterize the content of documents. WWW indices, like {{\sc Yahoo!}}\ provide a huge hierarchy of categories (topics) that touch every aspect of human endeavors. Such topics can be used as descriptors, similarly to the way librarians use for example, the Library of Congress cataloging system to annotate and categorize books. In the course of investigating this idea, we address the problem of automatic categorization of webpages in the {{\sc Yahoo!}}\ directory. We use Telltale as our classifier; Telltale uses n-grams to compute the similarity between documents. We experiment with various types of descriptions for the {{\sc Yahoo!}}\ categories and the webpages to be categorized. Our findings suggest that the best results occur when using the very brief descriptions of the {{\sc Yahoo!}}\ categorized entries; these brief descriptions are provided either by the entries' submitters or by the {{\sc Yahoo!}}\ human indexers and accompany most {{\sc Yahoo!}}\-indexed entries.}, } @inProceedings{Lai01, author = {Kwok-Yin Lai and Wai Lam}, title = {Meta-learning Models for Automatic Textual Document Categorization}, booktitle = {Proceedings of PAKDD-01, 5th Pacific-Asia Conferenece on Knowledge Discovery and Data Mining}, editor = {David Cheung and Qing Li and Graham Williams}, year = {2001}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Hong Kong, CN}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2035}, pages = {78--89}, url = {http://link.springer.de/link/service/series/0558/papers/2035/20350078.pdf}, abstract = {We investigate two meta-model approaches for the task of automatic textual document categorization. The first approach is the linear combination approach. Based on the idea of distilling the characteristics of how we estimate the merits of each component algorithm, we propose three different strategies for the linear combination approach. The linear combination approach makes use of limited knowledge in the training document set. To address this limitation, we propose the second meta-model approach, called Meta-learning Using Document Feature characteristics (MUDOF), which employs a meta-learning phase using document feature characteristics. Document feature characteristics, derived from the training document set, capture some inherent properties of a particular category. Extensive experiments have been conducted on a real-world document collection and satisfactory performance is obtained.}, } @article{Lai02, author = {Yu-Sheng Lai and Chung-Hsien Wu}, title = {Meaningful term extraction and discriminative term selection in text categorization via unknown-word methodology}, journal = {ACM Transactions on Asian Language Information Processing}, year = {2002}, number = {1}, volume = {1}, pages = {34--64}, url = {http://doi.acm.org/10.1145/595576.595579}, abstract = {In this article, an approach based on unknown words is proposed for meaningful term extraction and discriminative term selection in text categorization. For meaningful term extraction, a phrase-like unit (PLU)-based likelihood ratio is proposed to estimate the likelihood that a word sequence is an unknown word. On the other hand, a discriminative measure is proposed for term selection and is combined with the PLU-based likelihood ratio to determine the text category. We conducted several experiments on a news corpus, called MSDN. The MSDN corpus is collected from an online news Website maintained by the Min-Sheng Daily News, Taiwan. The corpus contains 44,675 articles with over 35 million words. The experimental results show that the system using a simple classifier achieved 95.31\% accuracy. When using a state-of-the-art classifier, kNN, the average accuracy is 96.40\%, outperforming all the other systems evaluated on the same collection, including the traditional term-word by kNN (88.52\%); sleeping-experts (82.22\%); sparse phrase by four-word sleeping-experts (86.34\%); and Boolean combinations of words by RIPPER (87.54\%). A proposed purification process can effectively reduce the dimensionality of the feature space from 50,576 terms in the word-based approach to 19,865 terms in the unknown word-based approach. In addition, more than 80\% of automatically extracted terms are meaningful. Experiments also show that the proportion of meaningful terms extracted from training data is relative to the classification accuracy in outside testing.}, } @inProceedings{Lam01, author = {Wai Lam and Kwok-Yin Lai}, title = {A Meta-Learning Approach for Text Categorization}, booktitle = {Proceedings of SIGIR-01, 24th ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and David J. Harper and Donald H. Kraft and Justin Zobel}, publisher = {ACM Press, New York, US}, address = {New Orleans, US}, year = {2001}, pages = {303--309}, url = {http://portal.acm.org/citation.cfm?doid=383952.384011}, abstract = {We investigate a meta-model approach, called Meta-learning Using Document Feature characteristics (MUDOF), for the task of automatic textual document categorization. It employs a meta-learning phase using document feature characteristics. Document feature characteristics, derived from the training document set, capture some inherent category-specific properties of a particular category. Different from existing categorization methods, MUDOF can automatically recommend a suitable algorithm for each category based on the category-specific statistical characteristics. Hence, different algorithms may be employed for different categories. Experiments have been conducted on a real-world document collection demonstrating the effectiveness of our approach. The results confirm that our meta-model approach can exploit the advantage of its component algorithms, and demonstrate a better performance than existing algorithms.}, } @inProceedings{Lam97, author = {Wai Lam and Kon F. Low and Chao Y. Ho}, title = {Using a Bayesian Network Induction Approach for Text Categorization}, booktitle = {Proceedings of IJCAI-97, 15th International Joint Conference on Artificial Intelligence}, editor = {Martha E. Pollack}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, year = {1997}, address = {Nagoya, JP}, pages = {745--750}, url = {}, abstract = {We investigate Bayesian methods for automatic document categorization and develop a new approach to this problem. Our new approach is based on a Bayesian network induction which does not rely on some major assumptions found in a previous method using the Bayesian independence classifier approach. The design of the new approach as well as its justification are presented. Experiments were conducted using a large scale document collection from Reuters news articles. The results show that our approach outperformed the Bayesian independence classifier as measured by a metric that combines precision and recall measures.}, } @inProceedings{Lam98, author = {Wai Lam and Chao Y. Ho}, title = {Using a generalized instance set for automatic text categorization}, booktitle = {Proceedings of SIGIR-98, 21st ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and Alistair Moffat and Van Rijsbergen, Cornelis J. and Ross Wilkinson and Justin Zobel}, publisher = {ACM Press, New York, US}, year = {1998}, address = {Melbourne, AU}, pages = {81--89}, url = {http://www.acm.org/pubs/articles/proceedings/ir/290941/p81-lam/p81-lam.pdf}, abstract = {We investigate several recent approaches for text categorization under the framework of similarity-based learning. They include two families of text categorization techniques, namely the k-nearest neighbor (k-NN) algorithm and linear classifiers. After identifying the weakness and strength of each technique, we propose a new technique known as the generalized instance set (GIS) algorithm by unifying the strengths of LNN and linear classifiers and adapting to characteristics of text categorization problems. We also explore some variants of our GIS approach. We have implemented our GIS algorithm, the ExpNet algorithm, and some linear classifiers. Extensive experiments have been conducted on two common document corpora, namely the OHSUMED collection and the Reuters-21578 collection. The results show that our new approach outperforms the latest LNN approach and linear classifiers in all experiments.}, } @inProceedings{Lam99, author = {Savio L. Lam and Dik L. Lee}, title = {Feature Reduction for Neural Network Based Text Categorization}, booktitle = {Proceedings of DASFAA-99, 6th IEEE International Conference on Database Advanced Systems for Advanced Application}, editor = {Arbee L. Chen and Frederick H. Lochovsky}, publisher = {IEEE Computer Society Press, Los Alamitos, US}, year = {1999}, address = {Hsinchu, TW}, pages = {195--202}, url = {http://dlib.computer.org/conferen/dasfaa/0084/pdf/00840195.pdf}, abstract = {In a text categorization model using an artificial neural network as the text classifier scalability is poor if the neural network is trained using the raw feature space since textural data has a very high-dimension feature space. We proposed and compared four dimensionality reduction techniques to reduce the feature space into an input space of much lower dimension for the neural network classifier. To test the effectiveness of the proposed model, experiments were conducted using a subset of the Reuters-22173 test collection for text categorization. The results showed that the proposed model was able to achieve high categorization effectiveness as measured by precision and recall. Among the four dimensionality reduction techniques proposed, principal component analysis was found to be the most effective in reducing the dimensionality of the feature space.}, } @article{Lam99a, author = {Lam, Wai and Ruiz, Miguel E. and Srinivasan, Padmini}, title = {Automatic text categorization and its applications to text retrieval}, journal = {IEEE Transactions on Knowledge and Data Engineering}, year = {1999}, number = {6}, volume = {11}, pages = {865--879}, url = {http://www.cs.uiowa.edu/~mruiz/papers/IEEE-TKDE.ps}, abstract = {We develop an automatic text categorization approach and investigate its application to text retrieval. The categorization approach is derived from a combination of a learning paradigm known as instance-based learning and an advanced document retrieval technique known as retrieval feedback. We demonstrate the effectiveness of our categorization approach using two real-world document collections from the MEDLINE database. Next, we investigate the application of automatic categorization to text retrieval. Our experiments clearly indicate that automatic categorization improves the retrieval performance compared with no categorization. We also demonstrate that the retrieval performance using automatic categorization achieves the same retrieval quality as the performance using manual categorization. Furthermore, detailed analysis of the retrieval performance on each individual test query is provided.}, } @inProceedings{Lang95, author = {Ken Lang}, title = {{\sc NewsWeeder}: learning to filter netnews}, booktitle = {Proceedings of ICML-95, 12th International Conference on Machine Learning}, editor = {Armand Prieditis and Stuart J. Russell}, address = {Lake Tahoe, US}, pages = {331--339}, year = {1995}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {}, abstract = {}, } @inProceedings{Lanquillon00, author = {Carsten Lanquillon}, title = {Learning from Labeled and Unlabeled Documents: A Comparative Study on Semi-Supervised Text Classification}, booktitle = {Proceedings of PKDD-00, 4th European Conference on Principles of Data Mining and Knowledge Discovery}, editor = {Djamel A. Zighed and Henryk Jan Komorowski and Jan M. Zytkow}, address = {Lyon, FR}, pages = {490--497}, year = {2000}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1910}, url = {http://link.springer.de/link/service/series/0558/papers/1910/19100490.pdf}, abstract = {Supervised learning algorithms usually require large amounts of training data to learn reasonably accurate classifiers. Yet, for many text classification tasks, providing labeled training documents is expensive, while unlabeled documents are readily available in large quantities. Learning from both, labeled and unlabeled documents, in a semi-supervised framework is a promising approach to reduce the need for labeled training documents. This paper compares three commonly applied text classifiers in the light of semi-supervised learning, namely a linear support vector machine, a similarity-based tfidf and a Naive Bayes classifier. Results on a real-world text datasets show that these learners may substantially benefit from using a large amount of unlabeled documents in addition to some labeled documents.}, } @inProceedings{Larkey96, author = {Leah S. Larkey and W. Bruce Croft}, title = {Combining classifiers in text categorization}, booktitle = {Proceedings of SIGIR-96, 19th ACM International Conference on Research and Development in Information Retrieval}, editor = {Hans-Peter Frei and Donna Harman and Peter Sch{\"{a}}uble and Ross Wilkinson}, publisher = {ACM Press, New York, US}, year = {1996}, address = {Z{\"{u}}rich, CH}, pages = {289--297}, url = {http://cobar.cs.umass.edu/pubfiles/1combo.ps.gz}, abstract = {Three different types of classifiers were investigated in the context of a text categorization problem in the medical domain: the automatic assignment of ICD9 codes to dictated inpatient discharge summaries. K-nearest-neighbour, relevance feedback, and Bayesian independence classifiers were applied individually and in combination. A combination of different classifiers produced better results than any single type of classifier. For this specific medical categorization problem, new query formulation and weighting methods used in the k-nearest-neighbor classifier improved performance.}, } @inProceedings{Larkey98, author = {Leah S. Larkey}, title = {Automatic essay grading using text categorization techniques}, booktitle = {Proceedings of SIGIR-98, 21st ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and Alistair Moffat and Van Rijsbergen, Cornelis J. and Ross Wilkinson and Justin Zobel}, publisher = {ACM Press, New York, US}, year = {1998}, address = {Melbourne, AU}, pages = {90--95}, url = {http://cobar.cs.umass.edu/pubfiles/ir-121.ps}, abstract = {Several standard text-categorization techniques were applied to the problem of automated essay grading. Bayesian independence classifiers and k-nearest-neighbor classifiers were trained to assign scores to manually-graded essays. These scores were combined with several other summary text measures using linear regression. The classifiers and regression equations were then applied to a new set of essays. The classifiers worked very well. The agreement between the automated grader and the final manual grade was as good as the agreement between human graders.}, } @inProceedings{Larkey99, author = {Leah S. Larkey}, title = {A patent search and classification system}, booktitle = {Proceedings of DL-99, 4th ACM Conference on Digital Libraries}, editor = {Edward A. Fox and Neil Rowe}, publisher = {ACM Press, New York, US}, year = {1999}, address = {Berkeley, US}, pages = {179--187}, url = {http://cobar.cs.umass.edu/pubfiles/ir-162.ps}, abstract = {We present a system for searching and classifying U.S. patent documents, based on Inquery. Patents are distributed through hundreds of collections, divided up by general area. The system selects the best collections for the query. Users can search for pants or classify patent text. The user interface helps users search in fields without requiring the knowledge of Inquery query operators. The system includes a unique phrase help facility, which helps users find and add phrases and terms related to those in their query.}, } @inProceedings{Lee00, author = {Hahn-Ming Lee and Chih-Ming Chen and Cheng-Wei Hwang}, title = {A neural network document classifier with linguistic feature selection}, booktitle = {Proceedings of IEA/AIE-00, 13th International Conference on Industrial and Engineering Applications of Artificial Intelligence and Expert Systems}, publisher = {}, editor = {}, year = {2003}, address = {New Orleans, US}, pages = {555--560}, url = {}, abstract = {}, } @inProceedings{Lee02, author = {Yong-Bae Lee and Sung H. Myaeng}, title = {Text Genre Classification with Genre-Revealing and Subject-Revealing Features}, booktitle = {Proceedings of SIGIR-02, 25th ACM International Conference on Research and Development in Information Retrieval}, editor = {Micheline Beaulieu and Ricardo Baeza-Yates and Sung Hyon Myaeng and Kalervo J{\"{a}}rvelin}, publisher = {ACM Press, New York, US}, address = {Tampere, FI}, year = {2002}, pages = {145--150}, url = {http://doi.acm.org/10.1145/564376.564403}, abstract = {Subject or prepositional content has been the focus of most classification research. Genre or style, on the other hand, is a different and important property of text, and automatic text genre classification is becoming important for classification and retrieval purposes as well as for some natural language processing research. In this paper, we present a method for automatic genre classification that is based on statistically selected features obtained from both subject-classified and genre classified training data. The experimental results show that the proposed method outperforms a direct application of a statistical learner often used for subject classification. We also observe that the deviation formula and discrimination formula using document frequency ratios also work as expected. We conjecture that this dual feature set approach can be generalized to improve the performance of subject classification as well.}, } @inProceedings{Lee02a, author = {Michael D. Lee}, title = {Fast Text Classification Using Sequential Sampling Processes}, booktitle = {Proceedings of the 14th Australian Joint Conference on Artificial Intelligence}, editor = {Markus Stumptner and Dan Corbett and Michael J. Brooks}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Adelaide, AU}, year = {2002}, pages = {309--320}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2256}, url = {http://link.springer.de/link/service/series/0558/papers/2256/22560309.pdf}, abstract = {A central problem in information retrieval is the automated classification of text documents. While many existing methods achieve good levels of performance, they generally require levels of computation that prevent them from making sufficiently fast decisions in some applied setting. Using insights gained from examining the way humans make fast decisions when classifying text documents, two new text classification algorithms are developed based on sequential sampling processes. These algorithms make extremely fast decisions, because they need to examine only a small number of words in each text document. Evaluation against the Reuters-21578 collection shows both techniques have levels of performance that approach benchmark methods, and the ability of one of the classifiers to produce realistic measures of confidence in its decisions is shown to be useful for prioritizing relevant documents.}, } @inProceedings{Lee02c, author = {Kang Hyuk Lee and Judy Kay and Byeong Ho Kang and Uwe Rosebrock}, title = {A Comparative Study on Statistical Machine Learning Algorithms and Thresholding Strategies for Automatic Text Categorization}, booktitle = {Proceedings of PRICAI-02, 7th Pacific Rim International Conference on Artificial Intelligence}, editor = {Mitsuru Ishizuka and Abdul Sattar}, publisher = {Springer Verlag, Heidelberg, DE}, address = {Tokyo, JP}, year = {2002}, pages = {444--453}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 2417}, url = {http://link.springer.de/link/service/series/0558/papers/2417/24170444.pdf}, abstract = {Two main research areas in statistical text categorization are similarity-based learning algorithms and associated thresholding strategies. The combination of these techniques significantly influences the overall performance of text categorization. After investigating two similarity-based classifiers (k-NN and Rocchio) and three common thresholding techniques (RCut, PCut, and SCut), we describe a new learning algorithm known as the keyword association network (KAN) and a new thresholding strategy (RinSCut) to improve performance over existing techniques. Extensive experiments have been conducted on the Reuters-21578 and 20-Newsgroups data sets. The experimental results show that our new approaches give better results for both micro-averaged F1 and macro-averaged F1 scores.}, } @article{Lehnert94, author = {Wendy Lehnert and Stephen Soderland and David Aronow and Fangfang Feng and Avinoam Shmueli}, title = {Inductive text classification for medical applications}, journal = {Journal of Experimental and Theoretical Artificial Intelligence}, year = {1994}, number = {1}, volume = {7}, pages = {49--80}, url = {}, abstract = {}, } @article{Leopold02, author = {Leopold, Edda and Kindermann, J{\"{o}}rg}, title = {Text Categorization with Support Vector Machines: How to Represent Texts in Input Space?}, journal = {Machine Learning}, year = {2002}, volume = {46}, number = {1/3}, pages = {423--444}, url = {http://www.wkap.nl/article.pdf?380516}, abstract = {The choice of the kernel function is crucial to most applications of support vector machines. In this paper, however, we show that in the case of text classification, term-frequency transformations have a larger impact on the performance of SVM than the kernel itself. We discuss the role of importance-weights (e.g. document frequency and redundancy), which is not yet fully understood in the light of model complexity and calculation cost, and we show that time consuming lemmatization or stemming can be avoided even when classifying a highly inflectional language like German.}, } @article{Lertnattee04, author = {Verayuth Lertnattee and Thanaruk Theeramunkong}, title = {Effect of term distributions on centroid-based text categorization}, journal = {Information Sciences}, year = {2004}, number = {1}, volume = {158}, pages = {89--115}, url = {http://dx.doi.org/10.1016/j.ins.2003.07.007}, abstract = {Most of traditional text categorization approaches utilize term frequency (tf) and inverse document frequency (idf) for representing importance of words and/or terms in classifying a text document. This paper describes an approach to apply term distributions, in addition to tf and idf, to improve performance of centroid-based text categorization. Three types of term distributions, called inter-class, intra-class and in-collection distributions, are introduced. These distributions are useful to increase classification accuracy by exploiting information of (1) term distribution among classes, (2) term distribution within a class and (3) term distribution in the whole collection of training data. In addition, this paper investigates how these term distributions contribute to weight each term in documents, e.g., a high term distribution of a word promotes or demotes importance or classification power of that word. To this end, several centroid-based classifiers are constructed with different term weightings. Using various data sets, their performances are investigated and compared to a standard centroid-based classifier (TDIDF) and a centroid-based classifier modified with information gain. Moreover, we also compare them to two well-known methods: k-NN and naive Bayes. In addition to a unigram model of document representation, a bigram model is also explored. Finally, the effectiveness of term distributions to improve classification accuracy is explored with regard to the training set size and the number of classes.}, } @article{Leung97, author = {Chi-Hong Leung and Wing-Kay Kan}, title = {A Statistical Learning Approach to Automatic Indexing of Controlled Index Terms}, journal = {Journal of the American Society for Information Science}, year = {1997}, number = {1}, pages = {55--67}, volume = {48}, url = {http://www3.interscience.wiley.com/cgi-bin/fulltext?ID=39602&PLACEBO=IE.pdf}, abstract = {A statistical learning approach to assigning controlled index terms is presented. In this approach, there are two processes: (1) The learning process and (2) the indexing process. The learning process constructs a relationship between an index term and the words relevant and irrelevant to it, based on the positive training set and negative training set, which are sample documents indexed by the index term, and those not indexed by it, respectively. The indexing process determines whether an index term is assigned to a certain document, based on the relationship constructed by the learning process, and the text found in the document. Furthermore, a learning feedback technique is introduced. This technique used in the learning process modifies the relationship between an index term and its relevant and irrelevant words to improve the learning performance and, thus, the indexing performance. Experimental results have shown that the statistical learning approach and the learning feedback technique are practical means to automatic indexing of controlled index terms.}, } @inProceedings{Lewis00, author = {Lewis, David D.}, title = {Machine learning for text categorization: background and characteristics}, booktitle = {Proceedings of the 21st Annual National Online Meeting}, editor = {Williams, Martha E.}, publisher = {Information Today, Medford, USA}, address = {New York, US}, year = {2000}, pages = {221--226}, url = {}, abstract = {Text categorization is of increasing interest in both controlled vocabulary indexing and other applications. Machine learning methods for automatically producing categorization rules have similarly seen increased attention, as a way to reduce the cost of fielding categorization systems. While the experimental literature on text categorization emphasizes effectiveness comparisons, we list a variety of other characteristics of learning approaches that are equally important to consider. Research on machine learning for text categorization, already advancing at a rapid pace, could be further accelerated if better test collections were available.}, } @article{Lewis04, author = {Lewis, David D. and Fan Li and Tony Rose and Yiming Yang}, title = {{Reuters Corpus Volume 1} as a text categorization test collection}, journal = {Journal of Machine Learning Research}, volume = {5}, month = {April}, pages = {361--397}, year = {2004}, url = {http://www.jmlr.org/papers/volume5/lewis04a/lewis04a.pdf}, abstract = {Reuters Corpus Volume I (RCV1) is an archive of over 800,000 manually categorized newswire stories recently made available by Reuters, Ltd. for research purposes. Use of this data for research on text categorization requires a detailed understanding of the real world constraints under which the data was produced. Drawing on interviews with Reuters personnel and access to Reuters documentation, we describe the coding policy and quality control procedures used in producing the RCV1 data, the intended semantics of the hierarchical category taxonomies, and the corrections necessary to remove errorful data. We refer to the original data as RCV1-v1, and the corrected data as RCV1-v2. We benchmark several widely used supervised learning methods on RCV1-v2, illustrating the collection's properties, suggesting new directions for research, and providing baseline results for future studies. We make available detailed, per-category experimental results, as well as corrected versions of the category assignments and taxonomy structures, via online appendices.}, } @inProceedings{Lewis91, author = {Lewis, David D.}, title = {Data extraction as text categorization: An experiment with the {MUC-3} corpus.}, booktitle = {Proceedings of MUC-3, 3rd Message Understanding Conference}, editor = {}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, address = {San Diego, US}, pages = {245--255}, year = {1991}, url = {http://www.research.att.com/~lewis/papers/lewis91c.ps}, abstract = {[no abstract]}, } @inProceedings{Lewis92, author = {Lewis, David D.}, title = {An evaluation of phrasal and clustered representations on a text categorization task}, booktitle = {Proceedings of SIGIR-92, 15th ACM International Conference on Research and Development in Information Retrieval}, editor = {Nicholas J. Belkin and Peter Ingwersen and Annelise Mark Pejtersen}, publisher = {ACM Press, New York, US}, address = {Kobenhavn, DK}, pages = {37--50}, year = {1992}, url = {http://www.research.att.com/~lewis/papers/lewis92b.ps}, abstract = {Syntactic phrase indexing and term clustering have been widely explored as text representation techniques for text retrieval. In this paper, we study the properties of phrasal and clustered indexing languages on a text categorization task, enabling us to study their properties in isolation from query interpretation issues. We show that optimal effectiveness occurs when using only a small proportion of the indexing terms available, and that effectiveness peaks at a higher feature set size and lower effectiveness level for a syntactic phrase indexing than for word-based indexing. We also present results suggesting that traditional term clustering methods are unlikely to provide significantly improved text representations. An improved probabilistic text categorization method is also presented.}, } @phdThesis{Lewis92a, author = {Lewis, David D.}, title = {Representation and learning in information retrieval}, school = {Department of Computer Science, University of Massachusetts}, address = {Amherst, US}, year = {1992}, url = {http://www.research.att.com/~lewis/papers/lewis91d.ps}, abstract = {This dissertation introduces a new theoretical model for text classification systems, including systems for document retrieval, automated indexing, electronic mail filtering, and similar tasks. The Concept Learning model emphasizes the role manual and automated feature selection and classifier formation in text classification. It enables drawing on results from statistics and machine learning in explaining the effectiveness of alternate representations of text, and specifies desirable characteristics of text representations. The use of syntactic parsing to produce indexing phrases has been widely investigated as a possible route to better text representations. Experiments with syntactic phrase indexing, however, have never yielded significant improvements in text retrieval performance. The Concept Learning model suggests that the poor statistical characteristics of a syntactic indexing phrase representation negate its dsirable semantic characteristics. The application of term clustering to this representation to improve its statistical properties while retaining its desirable meaning properties is proposed. Standard term clustering strategies from information retrieval (IR), based on cooccurence of indexing terms in documents or groups of documents, were tested on a syntactic indexing phrase representation. In experiments using a standard text retrieval test collection, small effectiveness improvements were obtained. As a means of evaluating representation quality, a text retrieval test collection introduces a number of confounding factors. In contrast, the text categorization task allows much cleaner determination of text representation properties. In preparation for the use of text categorization to study text representation, a more effective and theoretically well-founded probablistic text categorization algorithm was developed, building on work by Maron, Fuhr, and others. Text categorization experiments supported a number of predictions of the Concept Learning model about properties of phrasal representations, including dimensionality properties not previously measured for text representations. However, in carefully controlled experiments using syntactic phrases produced by Church's stochastic bracketer, in conjunction with reciprocal nearest neighbor clustering, term clustering was found to produce essentially no improvement in the properties of the phrasal representation. New cluster analysis approaches are proposed to remedy the problems found in traditional term clustering methods.}, } @inProceedings{Lewis94, author = {Lewis, David D. and Marc Ringuette}, title = {A comparison of two learning algorithms for text categorization}, booktitle = {Proceedings of SDAIR-94, 3rd Annual Symposium on Document Analysis and Information Retrieval}, publisher = {}, editor = {}, year = {1994}, address = {Las Vegas, US}, pages = {81--93}, url = {http://www.research.att.com/~lewis/papers/lewis94b.ps}, abstract = {This paper examines the use of inductive learning to categorize natural language documents into predefined content categories. Categorization of text is of increasing importance in information retrieval and natural language processing systems. Previous research on automated text categorization has mixed machine learning and knowledge engineering methods, making it difficult to draw conclusions about the performance of particular methods. In this paper we present empirical results on the performance of a Bayesian classifier and a decision tree learning algorithm on two text categorization data sets. We find that both algorithms achieve reasonable performance and allow controlled tradeoffs between false positives and false negatives. The stepwise feature selection in the decision tree algorithm is particularly effective in dealing with the large feature sets common in text categorization. However, even this algorithm is aided by an initial prefiltering of features, confirming the results found by Almuallim and Dietterich on artificial data sets. We also demonstrate the impact of the time-varying nature of category definitions.}, } @inProceedings{Lewis94a, author = {Lewis, David D. and Gale, William A.}, title = {A sequential algorithm for training text classifiers}, booktitle = {Proceedings of SIGIR-94, 17th ACM International Conference on Research and Development in Information Retrieval}, editor = {W. Bruce Croft and Van Rijsbergen, Cornelis J.}, publisher = {Springer Verlag, Heidelberg, DE}, year = {1994}, address = {Dublin, IE}, pages = {3--12}, note = {See also~\cite{Lewis95a}}, url = {http://www.research.att.com/~lewis/papers/lewis94c.ps}, abstract = {The ability to cheaply train text classifiers is critical to their use in information retrieval, content analysis, natural language processing, and other tasks involving data which is partly or fully textual. An algorithm for sequential sampling during machine learning of statistical classifiers was developed and tested on a newswire text categorization task. This method, which we call uncertainty sampling, reduced by as much as 500-fold the amount of training data that would have to be manually classified to achieve a given level of effectiveness.}, } @article{Lewis94b, author = {Lewis, David D. and Philip J. Hayes}, title = {Guest editors' introduction to the special issue on text categorization}, journal = {ACM Transactions on Information Systems}, volume = {12}, number = {3}, pages = {231}, year = {1994}, } @inProceedings{Lewis94c, author = {Lewis, David D. and Jason Catlett}, title = {Heterogeneous uncertainty sampling for supervised learning}, booktitle = {Proceedings of ICML-94, 11th International Conference on Machine Learning}, editor = {William W. Cohen and Haym Hirsh}, year = {1994}, address = {New Brunswick, US}, pages = {148--156}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.research.att.com/~lewis/papers/lewis94e.ps}, abstract = {Uncertainty sampling methods iteratively request class labels for training instances whose classes are uncertain despite the previous labeled instances. These methods can greatly reduce the number of instances that an expert need label. One problem with this approach is that the classifier best suited for an application may be too expensive to train or use during the selection of instances. We test the use of one classifier (a highly efficient probabilistic one) to select examples for training another (the C4.5 rule induction program). Despite being chosen by this heterogeneous approach, the uncertainty samples yielded classifiers with lower error rates than random samples ten times larger.}, } @inProceedings{Lewis95, author = {Lewis, David D.}, title = {Evaluating and optmizing autonomous text classification systems}, booktitle = {Proceedings of SIGIR-95, 18th ACM International Conference on Research and Development in Information Retrieval}, editor = {Edward A. Fox and Peter Ingwersen and Raya Fidel}, publisher = {ACM Press, New York, US}, year = {1995}, address = {Seattle, US}, pages = {246--254}, url = {http://www.research.att.com/~lewis/papers/lewis95b.ps}, abstract = {Text retrieval systems typically produce a ranking of documents and let a user decide how far down that ranking to go. In contrast, programs that filter text streams, software that categorizes documents, agents which alert users, and many other IR systems must make decisions without human input or supervision. It is important to define what constitutes good effectiveness for these autonomous systems, tune the systems to achieve the highest possible effectiveness, and estimate how the effectiveness changes as new data is processed. We show how to do this for binary text classification systems, emphasizing that different goals for the system lead to different optimal behaviors. Optimizing and estimating effectiveness is greatly aided if classifiers that explicitly estimate the probability of class membership are used.}, } @article{Lewis95a, author = {Lewis, David D.}, title = {A sequential algorithm for training text classifiers: corrigendum and additional data}, journal = {SIGIR Forum}, year = {1995}, pages = {13--19}, volume = {29}, number = {2}, url = {http://www.research.att.com/~lewis/papers/lewis95g.ps}, abstract = {Previously I compared the effectiveness of uncertainty sampling with that of random sampling and relevance sampling in choosing training data for a text categorization data set (Lewis and Gale, 1994). (Relevance sampling is the application of relevance feedback to producing a training sample.) I have discovered a bug in my experimental software which caused the relevance sampling results reported in the paper to be incorrect. (The uncertainty sampling and random sampling results in that paper were correct.) I have since fixed the bug and rerun the experiments. This note presents the corrected results, along with additional data supporting the original claim that uncertainty sampling has an advantage over relevance sampling in most training situations.}, } @inProceedings{Lewis95b, author = {David D. Lewis}, title = {The {TREC-4} filtering track: description and analysis}, booktitle = {Proceedings of TREC-4, 4th Text Retrieval Conference}, publisher = {National Institute of Standards and Technology, Gaithersburg, US}, editor = {Donna K. Harman and Ellen M. Voorhees}, year = {1995}, address = {Gaithersburg, US}, pages = {165--180}, url = {http://www.research.att.com/~lewis/papers/lewis96b.ps}, abstract = {The TREC-4 (4th Text REtrieval Conference) filtering track was an experiment in the evaluation of binary text classification systems. In contrast to ranking systems, binary text classification systems may need to produce result sets of any size, requiring that sampling be used to estimate their effectiveness. We present an effectiveness measure based on utility, and two sampling strategies (pooling and stratified sampling) for estimating the utility of the submitted sets. An evaluation of four sites was successfully carried out using this approach.}, } @inProceedings{Lewis96, author = {Lewis, David D. and Robert E. Schapire and James P. Callan and Ron Papka}, title = {Training algorithms for linear text classifiers}, booktitle = {Proceedings of SIGIR-96, 19th ACM International Conference on Research and Development in Information Retrieval}, editor = {Hans-Peter Frei and Donna Harman and Peter Sch{\"{a}}uble and Ross Wilkinson}, publisher = {ACM Press, New York, US}, year = {1996}, address = {Z{\"{u}}rich, CH}, pages = {298--306}, url = {http://www.research.att.com/~lewis/papers/lewis96d.ps}, abstract = {Systems for text retrieval, routing, categorization and other IR tasks rely heavily on linear classifiers. We propose that two machine learning algorithms, the Widrow-Hoff and EG algorithms, be used in training linear text classifiers. In contrast to most IR methods, theoretical analysis provides performance guarantees and guidance on parameter settings for these algorithms. Experimental data is presented showing Widrow-Hoff and EG to be more effective than the widely used Rocchio algorithm on several categorization and routing tasks.}, } @misc{Lewis97a, author = {Lewis, David D.}, title = {Reuters-21578 text categorization test collection. Distribution 1.0}, year = {1997}, note = {Available as {\tt http://www.daviddlewis.com/resources/testcollections/reuters21578/readme.txt}}, url = {http://www.daviddlewis.com/resources/testcollections/reuters21578/readme.txt}, abstract = {[no abstract]}, } @inProceedings{Lewis98, author = {Lewis, David D.}, title = {Naive (Bayes) at forty: The independence assumption in information retrieval.}, booktitle = {Proceedings of ECML-98, 10th European Conference on Machine Learning}, publisher = {Springer Verlag, Heidelberg, DE}, note = {Published in the ``Lecture Notes in Computer Science'' series, number 1398}, editor = {Claire N{\'{e}}dellec and C{\'{e}}line Rouveirol}, address = {Chemnitz, DE}, pages = {4--15}, year = {1998}, url = {http://www.research.att.com/~lewis/papers/lewis98b.ps}, abstract = {The naive Bayes classifier, currently experiencing a renaissance in machine learning, has long been a core technique in information retrieval. We review some of the variations of naive Bayes models used for text retrieval and classification, focusing on the distributional assumptions made about word occurrences in documents.}, } @inProceedings{Lewis99, author = {Lewis, David D. and Daniel L. Stern and Amit Singhal}, title = {{\sc Attics}: a software platform for on-line text classification}, booktitle = {Proceedings of SIGIR-99, 22nd ACM International Conference on Research and Development in Information Retrieval}, editor = {Marti A. Hearst and Fredric Gey and Richard Tong}, publisher = {ACM Press, New York, US}, address = {Berkeley, US}, year = {1999}, pages = {267--268}, url = {http://www.acm.org/pubs/articles/proceedings/ir/312624/p267-lewis/p267-lewis.pdf}, abstract = {Numerous systems for ranked retrieval on text databases have been implemented by both information retrieval researchers and in the commercial sector. In contrast, software for text categorization, message filtering, textual data mining, and related tasks is less common. ATTICS is an extensible text classification system we have implemented in C++. It supports incremental training and online application of classifiers and predictive models to streams of textual, numeric, symbolic, and hybrid data records. An object-oriented design allows easy addition of new preprocessors, machine learning algorithms, and classifier types.}, } @article{Li02, author = {Hang Li and Kenji Yamanishi}, title = {Text classification using ESC-based stochastic decision lists}, journal = {Information Processing and Management}, pages = {343--361}, year = {2002}, number = {3}, volume = {38}, url = {}, abstract = {We propose a new method of text classification using stochastic decision lists. A stochastic decision list is an ordered sequence of IF-THEN-ELSE rules, and our method can be viewed as a rule-based method for text classification having advantages of readability and refinability of acquired knowledge. Our method is unique in that decision lists are automatically constructed on the basis of the principle of minimizing extended stochastic complexity (ESC), and with it we are able to construct decision lists that have fewer errors in classification. The accuracy of classification achieved with our method appears better than or comparable to those of existing rule-based methods. We have empirically demonstrated that rule-based methods like ours result in high classification accuracy when the categories to which texts are to be assigned are relatively specific ones and when the texts tend to be short. We have also empirically verified the advantages of rule-based methods over non-rule-based ones.}, } @inProceedings{Li02a, author = {Xin Li and Dan Roth}, title = {Learning question classifiers}, booktitle = {Proceedings of COLING-02, 19th International Conference on Computational Linguistics}, editor = {}, publisher = {}, address = {Taipei, TW}, url = {http://l2r.cs.uiuc.edu/~danr/Papers/qc-coling02.pdf}, year = {2002}, abstract = {In order to respond correctly to a free form factual question given a large collection of texts, one needs to understand the question to a level that allows determining some of the constraints the question imposes on a possible answer. These constraints may include a semantic classification of the sought after answer and may even suggest using different strategies when looking for and verifying a candidate answer. This paper presents a machine learning approach to question classification. We learn a hierarchical classifier that is guided by a layered semantic hierarchy of answer types, and eventually classifies questions into finegrained classes. We show accurate results on a large collection of free-form questions used in TREC 10.}, } @inProceedings{Li03, author = {Cong Li and Ji-Rong Wen and Hang Li}, title = {Text Classification Using Stochastic Keyword Generation}, booktitle = {Proceedings of ICML-03, 20th International Conference on Machine Learning}, editor = {}, year = {2003}, address = {Washington, DC}, pages = {}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {}, abstract = {}, } @inProceedings{Li03a, author = {Fan Li and Yiming Yang}, title = {A Loss Function Analysis for Classification Methods in Text Categorization}, booktitle = {Proceedings of ICML-03, 20th International Conference on Machine Learning}, editor = {}, year = {2003}, address = {Washington, DC}, pages = {}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {}, abstract = {}, } @inProceedings{Li03b, author = {Tao Li and Shenghuo Zhu and Mitsunori Ogihara}, title = {Efficient multi-way text categorization via generalized discriminant analysis}, booktitle = {Proceedings of CIKM-03, 12th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {2003}, address = {New Orleans, US}, pages = {317--324}, url = {http://doi.acm.org/10.1145/956863.956924}, abstract = {Text categorization is an important research area and has been receiving much attention due to the growth of the on-line information and of Internet. Automated text categorization is generally cast as a multi-class classification problem. Much of previous work focused on binary document classification problems. Support vector machines (SVMs) excel in binary classification, but the elegant theory behind large-margin hyperplane cannot be easily extended to multi-class text classification. In addition, the training time and scaling are also important concerns. On the other hand, other techniques naturally extensible to handle multi-class classification are generally not as accurate as SVM. This paper presents a simple and efficient solution to multi-class text categorization. Classification problems are first formulated as optimization via discriminant analysis. Text categorization is then cast as the problem of finding coordinate transformations that reflects the inherent similarity from the data. While most of the previous approaches decompose a multi-class classification problem into multiple independent binary classification tasks, the proposed approach enables direct multi-class classification. By using Generalized Singular Value Decomposition (GSVD), a coordinate transformation that reflects the inherent class structure indicated by the generalized singular values is identified. Extensive experiments demonstrate the efficiency and effectiveness of the proposed approach.}, } @inProceedings{Li91, author = {Wei Li and B. Lee and F. Krausz and K. Sahin}, title = {Text classification by a neural network}, booktitle = {Proceedings of the 23rd Annual Summer Computer Simulation Conference}, editor = {}, publisher = {}, address = {Baltimore, US}, pages = {313--318}, year = {1991}, url = {}, abstract = {When banks process their free-form telex traffic, the first task is the classification of the telexes. Historically, several attempts have been made to automate this process, using various stock phrases as the features on which to base the classification. This is a problem in which there are large amounts of data available, but the rules for classification are not explicitly available. For solving these kinds of problems, neural networks have the advantage of extracting the underlying relationships between the input data and the output classes automatically. Based on this consideration, the authors have built a neural network classification system, which has three subsystems: a user-maintainable feature definition subsystem, a feature extraction subsystem, and a neural network subsystem. The neural network is simulated on a VAX computer with a fast learning algorithm, and is combined with some non-statistical knowledge from the feature definition system. Above 90\% correct recognition rates have been achieved for the major categories concerned. The system is also applicable to text classification problems other than telex classification.}, } @inProceedings{Li97, author = {Hang Li and Kenji Yamanishi}, title = {Document classification using a finite mixture model}, booktitle = {Proceedings of ACL-97, 35th Annual Meeting of the Association for Computational Linguistics}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, editor = {Philip R. Cohen and Wolfgang Wahlster}, year = {1997}, address = {Madrid, ES}, pages = {39--47}, url = {http://xxx.lanl.gov/ps/cmp-lg/9705005}, abstract = {We propose a new method of classifying documents into categories. The simple method of conducting hypothesis testing over word-based distributions in categories suffers from the data sparseness problem. In order to address this difficulty, Guthrie et.al. have developed a method using distributions based on hard clustering of words, i.e., in which a word is assigned to a single cluster and words in the same cluster are treated uniformly. This method might, however, degrade classification results, since the distributions it employs are not always precise enough for representing the differences between categories. We propose here the use of soft clustering of words, i.e., in which a word can be assigned to several different clusters and each cluster is characterized by a specific word probability distribution. We define for each document category a finite mixture model, which is a linear combination of the probability distributions of the clusters. We thereby treat the problem of classifying documents as that of conducting statistical hypothesis testing over finite mixture models. In order to accomplish this testing, we employ the EM algorithm which helps efficiently estimate parameters in a finite mixture model. Experimental results indicate that our method outperforms not only the method using distributions based on hard clustering, but also the method using word-based distributions and the method based on cosine-similarity.}, } @article{Li98a, author = {Li, Yong H. and Jain, Anil K.}, title = {Classification of text documents}, journal = {The Computer Journal}, year = {1998}, volume = {41}, number = {8}, pages = {537--546}, url = {}, abstract = {The exponential growth of the Internet has led to a great deal of interest in developing useful and efficient tools and software to assist users in searching the Web. Document retrieval, categorization, routing and filtering can all be formulated as classification problems. However, the complexity of natural languages and the extremely high dimensionality of the feature space of documents have made this classification problem very difficult. We investigate four different methods for document classification: the naive Bayes classifier, the nearest neighbour classifier, decision trees and a subspace method. These were applied to seven-class Yahoo news groups (business, entertainment, health, international, politics, sports and technology) individually and in combination, We studied three classifier combination approaches: simple voting, dynamic classifier selection and adaptive classifier combination. Our experimental results indicate that the naive Bayes classifier and the subspace method outperform the other two classifiers on our data sets. Combinations of multiple classifiers did not always improve the classification accuracy compared to the best individual classifier. Among the three different combination approaches, our adaptive classifier combination method introduced here performed the best.}, } @inProceedings{Li99, author = {Hang Li and Kenji Yamanishi}, title = {Text classification using ESC-based stochastic decision lists}, booktitle = {Proceedings of CIKM-99, 8th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {1999}, address = {Kansas City, US}, pages = {122--130}, url = {http://www.acm.org/pubs/articles/proceedings/cikm/319950/p122-li/p122-li.pdf}, abstract = {We propose a new method of text classification using stochastic decision lists. A stochastic decision list is an ordered sequence of IF-THEN rules, and our method can be viewed as a rule-based method for text clsssification having advantages of readability and refinability of acquired knowledge. Our method is unique in that decision lists are automatically constructed on the basis of the principle of minimizing Extended Stochastic Complexity (ESC), and with it we are able to construct decision lists that have fewer errors in classification. The accuracy of classification achieved with our method appears better than or comparable to those of existing rule-based methods.}, } @inProceedings{Liao02, author = {Yihua Liao and V. Rao Vemuri}, title = {Using Text Categorization Techniques for Intrusion Detection}, booktitle = {Proceedings of the 11th USENIX Security Symposium}, publisher = {}, editor = {Dan Boneh}, year = {2002}, address = {San Francisco, US}, pages = {51--59}, url = {http://www.usenix.org/publications/library/proceedings/sec02/liao.html}, abstract = {A new approach, based on the k-Nearest Neighbor (kNN) classifier, is used to classify program behavior as normal or intrusive. Short sequences of system calls have been used by others to characterize a program's normal behavior before. However, separate databases of short system call sequences have to be built for different programs, and learning program profiles involves time-consuming training and testing processes. With the kNN classifier, the frequencies of system calls are used to describe the program behavior. Text categorization techniques are adopted to convert each process to a vector and calculate the similarity between two program activities. Since there is no need to learn individual program profiles separately, the calculation involved is largely reduced. Preliminary experiments with 1998 DARPA BSM audit data show that the kNN classifier can effectively detect intrusive attacks and achieve a low false positive rate.}, } @article{Liddy94, author = {Elizabeth D. Liddy and Woojin Paik and Edmund S. Yu}, title = {Text categorization for multiple users based on semantic features from a machine-readable dictionary}, journal = {ACM Transactions on Information Systems}, year = {1994}, number = {3}, volume = {12}, pages = {278--295}, url = {http://www.acm.org/pubs/articles/journals/tois/1994-12-3/p278-liddy/p278-liddy.pdf}, abstract = {The text categorization module described in the paper provides a front-end filtering function for the larger DR-LINK text retrieval system (Liddy and Myaeng 1993). The module evaluates a large incoming stream of documents to determine which documents are sufficiently similar to a profile at the broad subject level to warrant more refined representation and matching. To accomplish this task, each substantive word in a text is first categorized using a feature set based on the semantic subject field codes (SFCs) assigned to individual word senses in a machine-readable dictionary. When tested on 50 user profiles and 550 megabytes of documents, results indicate that the feature set that is the basis of the text categorization module and the algorithm that establishes the boundary of categories of potentially relevant documents accomplish their tasks with a high level of performance. This means that the category of potentially relevant documents for most profiles would contain at least 80\% of all documents later determined to be relevant to the profile. The number of documents in this set would be uniquely determined by the system's category-boundary predictor, and this set is likely to contain less than 5\% of the incoming stream of documents.}, } @inProceedings{Liere97, author = {Ray Liere and Prasad Tadepalli}, title = {Active learning with committees for text categorization}, booktitle = {Proceedings of AAAI-97, 14th Conference of the American Association for Artificial Intelligence}, editor = {}, publisher = {AAAI Press, Menlo Park, US}, year = {1997}, pages = {591--596}, address = {Providence, US}, url = {http://www.rdrop.com/~lierer/aaai97.ps}, abstract = {In many real-world domains, supervised learning requires a large number of training examples. In this paper, we describe an active learning method that uses a committee of learners to reduce the number of training examples required for learning. Our approach is similar to the Query by Committee framework, where disagreement among the committee members on the predicted label for the input part of the example is used to signal the need for knowing the actual value of the label. Our experiments are conducted in the text categorization domain, which is characterized by a large number of features, many of which are irrelevant. We report here on experiments using a committee of Winnow-based learners and demonstrate that this approach can reduce the number of labeled training examples required over that used by a single Winnow learner by 1-2 orders of magnitude.}, } @inProceedings{Liere98, author = {Ray Liere and Prasad Tadepalli}, title = {Active Learning with Committees: Preliminary Results in Comparing Winnow and Perceptron in Text Categorization}, booktitle = {Proceedings of CONALD-98, 1st Conference on Automated Learning and Discovery}, editor = {}, publisher = {AAAI Press, Menlo Park, US}, year = {1998}, pages = {}, address = {Pittsburgh, US}, url = {http://www.rdrop.com/~lierer/conald98.ps}, abstract = {The availability of vast amounts of information on the World Wide Web has created a big demand for automatic tools to organize and index that information. Unfortunately, the paradigm of supervised machine learning is ill-suited to this task, as it assumes that the training examples are classified by a teacher - usually a human. In this paper, we describe an active learning method based on Query by Committee (QBC) that reduces the number of labeled training examples (text documents) required for learning by 1-2 orders of magnitude.}, } @inProceedings{Lim99, author = {Lim, Joo Hwee}, title = {Learnable visual keywords for image classification}, booktitle = {Proceedings of DL-99, 4th ACM Conference on Digital Libraries}, editor = {Edward A. Fox and Neil Rowe}, publisher = {ACM Press, New York, US}, year = {1999}, address = {Berkeley, US}, pages = {139--145}, url = {http://www.acm.org/pubs/articles/proceedings/dl/313238/p139-lim/p139-lim.pdf}, abstract = {Automatic categorization of multimedia documents is an important function for a digital library system. While text categorization has received much attentions by IR researchers, classification of visual data is at its infancy stage. In this paper, we propose a notion of visual keywords for similarity matching between visual contents. Visual keywords can be constructed automatically from samples of visual data through supervised/unsupervised learning. Given a visual content, the occurrences of visual keywords are detected, summarized spatially, and coded via singular value decomposition to arrive at a concise coded description. The methods to create, detect, summarize, select, and code visual keywords will be detailed. Last but not least, we describe an evaluation experiment that classifies professional nature scenery photographs to demonstrate the effectiveness and efficiency of visual keywords for automatic categorization of images in digital libraries.}, } @article{Liu01, author = {Zhi-Qiang Liu and Ya-Jun Zhang}, title = {A competitive neural network approach to web-page categorization}, journal = {International Journal of Uncertainty, Fuzziness and Knowledge-Based Systems}, volume = {9}, number = {6}, pages = {731--741}, year = {2001}, } @inProceedings{Liu02, author = {Yan Liu and Yiming Yang and Jaime Carbonell}, title = {Boosting to Correct the Inductive Bias for Text Classification}, booktitle = {Proceedings of CIKM-02, 11th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {2002}, address = {McLean, US}, pages = {348--355}, url = {http://doi.acm.org/10.1145/584792.584850}, abstract = {This paper studies the effects of boosting in the context of different classification methods for text categorization, including Decision Trees, Naive Bayes, Support Vector Machines (SVMs) and a Rocchio-style classifier. We identify the inductive biases of each classifier and explore how boosting, as an error-driven resampling mechanism, reacts to those biases. Our experiments on the Reuters-21578 benchmark show that boosting is not effective in improving the performance of the base classifiers on common categories. However, the effect of boosting for rare categories varies across classifiers: for SVMs and Decision Trees, we achieved a 13-17\% performance improvement in macro-averaged F1 measure, but did not obtain substantial improvement for the other two classifiers. This interesting finding of boosting on rare categories has not been reported before.}, } @inProceedings{Liu03, author = {Yan Liu and Jaime Carbonell and Rong Jin}, title = {A New Pairwise Ensemble Approach for Text Classification}, booktitle = {Proceedings of ECML-03, 14th European Conference on Machine Learning}, publisher = {}, editor = {}, year = {2003}, address = {Dubrovnik, HK}, pages = {}, url = {}, abstract = {}, } @article{Lodhi02, author = {Huma Lodhi and Craig Saunders and John Shawe-Taylor and Nello Cristianini and Chris Watkins}, title = {Text Classification using String Kernels}, journal = {Journal of Machine Learning Research}, volume = {2}, pages = {419--444}, year = {2002}, url = {http://www.ai.mit.edu/projects/jmlr/papers/volume2/lodhi02a/lodhi02a.pdf}, abstract = {We propose a novel approach for categorizing text documents based on the use of a special kernel. The kernel is an inner product in the feature space generated by all subsequences of length k. A subsequence is any ordered sequence of k characters occurring in the text though not necessarily contiguously. The subsequences are weighted by an exponentially decaying factor of their full length in the text, hence emphasising those occurrences that are close to contiguous. A direct computation of this feature vector would involve a prohibitive amount of computation even for modest values of k, since the dimension of the feature space grows exponentially with k. The paper describes how despite this fact the inner product can be efficiently evaluated by a dynamic programming technique. Experimental comparisons of the performance of the kernel compared with a standard word feature space kernel (Joachims, 1998) show positive results on modestly sized datasets. The case of contiguous subsequences is also considered for comparison with the subsequences kernel with different decay factors. For larger documents and datasets the paper introduces an approximation technique that is shown to deliver good approximations efficiently for large datasets.}, } @inProceedings{Macskassy01, author = {Sofus A. Macskassy and Haym Hirsh and Arunava Banerjee and Aynur A. Dayanik}, title = {Using Text Classifiers for Numerical Classification}, booktitle = {Proceeding of IJCAI-01, 17th International Joint Conference on Artificial Intelligence}, editor = {Bernhard Nebel}, address = {Seattle, US}, year = {2001}, pages = {885--890}, url = {http://www.cs.rutgers.edu/~sofmac/paper/ijcai2001/macskassy-ijcai2001.pdf}, abstract = {Consider a supervised learning problem in which examples contain both numerical- and text-valued features. To use traditional feature-vector- based learning methods, one could treat the presence or ab-sence of a word as a Boolean feature and use these binary-valued features together with the numerical features. However, the use of a text-classification system on this is a bit more problematic - in the most straight-forward approach each number would be considered a distinct token and treated as a word. This paper presents an alter-native approach for the use of text classification methods for super-vised learning problems with numerical-valued features in which the numerical features are converted into bag-of-words features, thereby making them directly usable by text classification methods. We show that even on purely numerical-valued data the results of text-classification on the derived text-like representation outperforms the more naive numbers-as-tokens representation and, more importantly, is competitive with mature numerical classification methods such as C4.5 and Ripper.}, } @article{Macskassy03, author = {Sofus A. Macskassy and Haym Hirsh and Arunava Banerjee and Aynur A. Dayanik}, title = {Converting numerical classification into text classification}, journal = {Artificial Intelligence}, volume = {143}, number = {1}, year = {2003}, pages = {51--77}, url = {}, abstract = {}, } @inProceedings{Macskassy03a, author = {Sofus A. Macskassy and Haym Hirsh}, title = {Adding numbers to text classification}, booktitle = {Proceedings of CIKM-03, 12th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {}, year = {2003}, address = {New Orleans, US}, pages = {240--246}, url = {http://doi.acm.org/10.1145/956863.956910}, abstract = {Many real-world problems involve a combination of both text- and numerical-valued features. For example, in email classification, it is possible to use instance representations that consider not only the text of each message, but also numerical-valued features such as the length of the message or the time of day at which it was sent. Text-classification methods have thus far not easily incorporated numerical features. In earlier work we described an approach for converting numerical features into bags of tokens so that text classification methods can be applied to numerical classification problems, and showed that the resulting learning methods are competitive with traditional numerical classification methods. In this paper we use this as a way to learn on problems that involve a combination of text and numbers. We show that the results outperform competing methods. Further, we show that selecting a best classification method using text-only features and then adding numerical features to the problem (as might happen if numerical features are only later added to a pre existing text-classification problem) gives performance that rivals a more time-consuming approach of evaluating all classification methods using the full set of both text and numerical features.}, } @article{Maderlechner97, author = {Maderlechner, G. and Suda, P. and Bruckner, T.}, title = {Classification of documents by form and content}, journal = {Pattern Recognition Letters}, pages = {1225--1231}, year = {1997}, volume = {18}, number = {11/13}, url = {}, abstract = {This paper presents a modular software system, which classifies a large variety of office documents according to layout form and textual content. It consists of the following components: layout analysis, pre-classification, OCR interface, fuzzy string matching, text categorization, lexical, syntactical and semantic analysis. The system has been applied to the following tasks: presorting of forms, reports and letters, index extraction for archiving and retrieval, page type classification and text column analysis of real estate register documents, in-house mail sorting and electronic distribution to departments. The architecture, modules, and practical results are described.}, } @article{Manevitz01, author = {Larry M. Manevitz and Malik Yousef}, title = {One-Class {SVMs} for Document Classification}, journal = {Journal of Machine Learning Research}, volume = {2}, month = {December}, pages = {139--154}, year = {2001}, url = {http://www.ai.mit.edu/projects/jmlr/papers/volume2/manevitz01a/manevitz01a.pdf}, abstract = {We implemented versions of the SVM appropriate for one-class classification in the context of information retrieval. The experiments were conducted on the standard Reuters data set. For the SVM implementation we used both a version of Schoelkopf et al. and a somewhat different version of one-class SVM based on identifying ``outlier" data as representative of the second-class. We report on experiments with different kernels for both of these implementations and with different representations of the data, including binary vectors, tf-idf representation and a modification called ``Hadamard" representation. Then we compared it with one-class versions of the algorithms prototype (Rocchio), nearest neighbor, naive Bayes, and finally a natural one-class neural network classification method based on ``bottleneck" compression generated filters. The SVM approach as represented by Schoelkopf was superior to all the methods except the neural network one, where it was, although occasionally worse, essentially comparable. However, the SVM methods turned out to be quite sensitive to the choice of representation and kernel in ways which are not well understood; therefore, for the time being leaving the neural network approach as the most robust.}, } @inBook{Manning99a, author = {Christopher Manning and Hinrich Sch{\"{u}}tze}, title = {Foundations of Statistical Natural Language Processing}, publisher = {The MIT Press}, address = {Cambridge, US}, year = {1999}, chapter = {16: Text Categorization}, pages = {575--608}, url = {}, abstract = {}, } @article{Maron61, author = {M.E. Maron}, title = {Automatic indexing: an experimental inquiry}, year = {1961}, journal = {Journal of the Association for Computing Machinery}, volume = {8}, number = {3}, pages = {404--417}, url = {http://www.acm.org/pubs/articles/journals/jacm/1961-8-3/p404-maron/p404-maron.pdf}, abstract = {This inquiry examines a technique for automatically classifying (indexing) documents according to their subject content. The task, in essence, is to have a computing machine read a document and on the basis of the occurrence of selected clue words decide to which of many subject categories the document in question belongs. This paper describes the design, execution and evaluation of a modest experimental study aimed at testing empirically one statistical technique for automatic indexing.}, } @inProceedings{Marton05, author = {Yuval Marton and Ning Wu and Lisa Hellerstein}, title = {On compression-based text classification}, booktitle = {Proceedings of ECIR-05, 27th European Conference on Information Retrieval}, publisher = {Springer Verlag}, editor = {David E. Losada and Juan M. Fern{'{a}}ndez-Luna}, address = {Santiago De Compostela, ES}, year = {2005}, pages = {300--314}, url = {}, abstract = {Compression-based text classification methods are easy to apply, requiring virtually no preprocessing of the data. Most such methods are character-based, and thus have the potential to automatically capture non-word features of a document, such as punctuation, word-stems, and features spanning more than one word. However, compression-based classification methods have drawbacks (such as slow running time), and not all such methods are equally effective. We present the results of a number of experiments designed to evaluate the effectiveness and behavior of different compression-based text classification methods on English text. Among our experiments are some specifically designed to test whether the ability to capture non-word features causes character-based text compression methods to achieve more accurate classification.}, } @inProceedings{Masand92, author = {Briji Masand and Gordon Linoff and David Waltz}, title = {Classifying news stories using memory-based reasoning}, booktitle = {Proceedings of SIGIR-92, 15th ACM International Conference on Research and Development in Information Retrieval}, editor = {Nicholas J. Belkin and Peter Ingwersen and Annelise Mark Pejtersen}, publisher = {ACM Press, New York, US}, address = {Kobenhavn, DK}, pages = {59--65}, year = {1992}, url = {http://www.acm.org/pubs/articles/proceedings/ir/133160/p59-masand/p59-masand.pdf}, abstract = {We describe a method for classifying news stories using Memory Based Reasoning (MBR) a k-nearest neighbor method), that does not require manual topic definitions. Using an already coded training database of about 50,000 stories from the Dow Jones Press Release News Wire, and SEEKER [Stanfill] (a text retrieval system that supports relevance feedback) as the underlying match engine, codes are assigned to new, unseen stories with a recall of about 80\% and precision of about 70\%. There are about 350 different codes to be assigned. Using a massively parallel supercomputer, we leverage the information already contained in the thousands of coded stories and are able to code a story in about 2 seconds. Given SEEKER, the text retrieval system, we achieved these results in about two person-months. We believe this approach is effective in reducing the development time to implement classification systems involving large number of topics for the purpose of classification, message routing etc.}, } @inCollection{Masand94, author = {Briji Masand}, title = {Optimising confidence of text classification by evolution of symbolic expressions}, booktitle = {Advances in genetic programming}, publisher = {The MIT Press}, address = {Cambridge, US}, year = {1994}, chapter = {21}, editor = {Kenneth E. Kinnear}, pages = {459--476}, url = {}, abstract = {}, } @inProceedings{Matsuda98, author = {Katsushi Matsuda and Toshikazu Fukushima}, title = {Task-oriented {World Wide Web} retrieval by document type classification}, booktitle = {Proceedings of CIKM-98, 7th ACM International Conference on Information and Knowledge Management}, publisher = {ACM Press, New York, US}, editor = {Georges Gardarin and James C. French and Niki Pissinou and Kia Makki and Luc Bouganim}, year = {1998}, address = {Bethesda, US}, pages = {109--113}, url = {http://www.acm.org/pubs/articles/proceedings/cikm/319950/p109-matsuda/p109-matsuda.pdf}, abstract = {This paper proposes a novel approach to accurately searching Web pages for relevant information in problem solving by specifying a Web document category instead of the user's task. Accessing information from World Wide Web pages as an approach to problem solving has become commonplace. However, such a search is difficult with current search services, since these services only provide keyword-based search methods that are equivalent to narrowing down the target references according to domains. However, problem solving usually involves both a domain and a task. Accordingly,' our approach is based on problem solving tasks. To specify a user's problem solving task, we introduce the concept of document types that directly relate to the problem solving tasks; with this approach, users can easily designate problem solving tasks. We implemented PageTypeSearch system based on our approach. Classifier of PageTypeSearch classifies Web pages into the document types by comparing their pages with typical structural characteristics of the types. We compare PageTypeSearch using the document type-indices with a conventional keyword-based search system in experiments. The average precision of the document type-based search is 88.9\%, while the average precision of the keyword-based search is 31.2\%. Moreover, the number of irrelevant references gathered by our system is about one-thirteenth that of traditional keyword-based search systems. Our approach has practical advantages for problem solving by introducing the viewpoint of tasks to achieve higher performance.}, } @inProceedings{McCallum98, author = {Andrew K. McCallum and Kamal Nigam}, title = {Employing {EM} in pool-based active learning for text classification}, booktitle = {Proceedings of ICML-98, 15th International Conference on Machine Learning}, editor = {Jude W. Shavlik}, year = {1998}, address = {Madison, US}, pages = {350--358}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.cs.cmu.edu/~mccallum/papers/emactive-icml98.ps.gz}, abstract = {The paper shows how a text classifier's need for labeled training documents can be reduced by taking advantage of a large pool of unlabeled documents. We modify the Query-by-Committee (QBC) method of active learning to use the unlabeled pool for explicitly estimating document density when selecting examples for labeling. Then active learning is combined with Expectation-Maximization in order to ``fill in'' the class labels of those documents that remain unlabeled. Experimental results show that the improvements to active learning require less than two-thirds as many labeled training examples as previous QBC approaches, and that the combination of EM and active learning requires only slightly more than half as many labeled training examples to achieve the same accuracy as either the improved active learning or EM alone.}, } @inProceedings{McCallum98b, author = {Andrew K. McCallum and Ronald Rosenfeld and Tom M. Mitchell and Andrew Y. Ng}, title = {Improving text classification by shrinkage in a hierarchy of classes}, booktitle = {Proceedings of ICML-98, 15th International Conference on Machine Learning}, editor = {Jude W. Shavlik}, year = {1998}, address = {Madison, US}, pages = {359--367}, publisher = {Morgan Kaufmann Publishers, San Francisco, US}, url = {http://www.cs.cmu.edu/~mccallum/papers/hier-icml98.ps.gz}, abstract = {When documents are organized in a large number of topic categories, the categories are often arranged in a hierarchy. The US pate