Type mismatches are a common situation in programming languages.
One of the solutions to this problem is type coercion: systematic
type changes
whenever such mismatches are detected. In this talk I show the
usefulness of type coercion for the formal semantics of plurals in
natural languages. A determiner word like all or three
basically
functions
as a relation between sets of atomic elements. A collective verb like
gather functions as a set of sets of elements. When a mismatch
between such items is detected (e.g. in the sentence all the students
gathered), semantic operations are defined to coerce a function ranging over
atoms into a function ranging over sets and vice versa. The formal
nature of this process is elaborated and its implications for the
study of plurals are discussed.
Logical entailment relations play a crucial role in the semantics of
natural language indicative sentences. Groenendijk and Stokhof (1997)
suggest that entailment relations should also be taken as driving the
semantics of questions. These relations include answerhood (the relation
between a question and its answers) and question entailment (similar to
the computer scientist's notion of reducing one problem to another). We
present a new logical interpretation of questions, based on these relations
and show that it is able to explain some notoriously difficult problems in
the semantics of questions. Our theory is based on interpreting sentences
over an algebraic structure called a bilattice originally used as a
framework for multi-valued reasoning in logic programming and AI
(Ginsberg 1988, Fitting 1991).
One of the approaches to natural language semantics
states that what we are interested in is an ability to explain and
predict entailments between sentences in natural language and not
only
the meaning of specific sentences per se. In his
PhD thesis, V. Sanchez proposes that certain kinds of entailments can be explained
by means of monotonicity based reasoning. Entailment between
sentences is viewed as a partial order relation that can be
calculated from orders between sub-sentential expressions.
This computation is based on the treatment of some natural language expressions
as monotonic functions.
We present a possible
formalization of Sanchez' work and develop an algorithm that
finds proofs of monotonicity based inferences, together
with a prototype of a system that computes inferences using this
algorithm.
This talk takes a modal look at predicate logic.
The first part of the talk is an independent
re-axiomatization of predicate logic in a
modal system which is one-to-one (up to alphabetic
variants) with predicate logic. The second part
of the talk explores the intuition that it is
the existence of a uniform base set for quantification
that makes predicate logic undecidable: a fragment
of the system in part one is presented, in which
the domain of quantification may grow in nested
quantification, and that is claimed to be decidable.
The talk will survey some theoretical background on the use of
economy principles in linguistics, within the Minimalist Program of Chomsky.
These principles will be demonstrated using a case study of
intonational stress shift. For instance: the informational difference
between sentences like (1) and (2) below.
In this talk I will present ongoing research on a new probabilistic model for parsing
natural language: the Tree-gram model. The model is computational and aims at capturing
actual (as opposed to idealized - as in Linguistics) human language behavior in a given
domain of language use through automatic supervised learning from a tree-bank. The probabilistic
component of the model addresses ambiguity resolution and is based on two principles:
1) probabilities are better when they are conditioned on actual words (i.e. lexicalization),
and 2) memorizing as many of the syntactic relations together with their (joint) probabilities
is better than assuming full independence. The first principle underlies all current models
that are based on Dependency-parsing, in particular models that are based on lexical-dependencies
(the UPenn school). The second principle underlies the Memory-Based approach represented by the
Data-Oriented Parsing (DOP) model (University of Amsterdam).
In this talk I will provide some evidence that the two approaches (UPenn and Amsterdam) have
complementary aspects of modeling. Subsequently, I will present the Tree-gram model which combines
complementary aspects from the two approaches. Also, I will review an empirical experiment with a
simplified version of the Tree-gram model that is based only on Data-Oriented Parsing. The experiment
is the first ever conducted with a DOP-like model on the (relatively large) Wall Street Journal corpus.
The result of this experiment is a surprise to people working only with lexical-dependencies: although
the simplified Tree-gram model does not condition all its probabilities on actual words, it does achieve
accuracy that is close to that achieved by some lexical-dependency models and is far better than
Probabilistic Context-Free Grammars. This implies that DOP-like syntactic relations are useful in parsing
and could improve models based on lexical-dependencies. Time permitting we will also discuss some problems
with The Tree-gram model and speculate on future solutions.
This talk concerns the semantics of temporal and spatial preposition
phrases in sentences which report the occurrence of events or the
holding of states:
In theoretical linguistics, causative and middle verbs are usually derived
by independent operations. But cross linguistically, both mark the same
transitivity alternation. The paper proposes a unified syntactic
system for the derivation of both types of verbs, which, moreover, sheds
new light on problems in the interface of semantics and morphology. One
problem is the impossibility, mostly ignored in linguistic theory, of
deriving the semantics of middle verbs from that of the corresponding
transitive verbs. The second is explaining the identity found cross
linguistically between middle and reflexive morphology. The third is
providing an alternative to the "event-decomposition" account of causative
verbs.
(1) My neighbour is building A DESK.
(2) My NEIGHBOUR is building a desk.
(capital letters designate intonation stress)
(1) Mary kissed John in the garden
(2) Mary is asleep in the garden.
(3) Mary was asleep in the garden at 2 o' clock.
In earlier work with Nissim Francez, a detailed semantics of temporal
preposition phrases is provided. The talk will briefly review this
earlier work and consider its extension to spatial prepositions. With
event-reporting sentences, all goes well: spatial preposition phrases
function so as to locate events in space just as their temporal
counterparts function so as to locate events in time. With
state-reporting sentences, by contrast, spatial preposition phrases
often function so as to replace the participating individuals by
those parts of them intersecting the region specified in the
complement of the preposition phrase. We draw some parallels between
the two ways in which spatial preposition phrases function and
speculate about the possibilities for a a unified account.
The paper develops a non-lexicalist account, based on the Semitic verbal
binyanim, of the semantics of causative and middle morphology. In the
Semitic languages, causative and middle verbs alike are derived from
roots by particular binyanim. The paper establishes that the Semitic
binyanim denote voice (of which middle is a possible value) and agency,
the thematic role of the verb's external argument (of which causative is
one possible value). According to the present analysis, this form-meaning
corresponddence is mediated by syntax, which allows the parallel
compositional construction of the form and the meaning of a verb from the
forms and the meanings of its root and binyan.
I take a root to denote either a property of events or a relation between
individuals and events. In the unmarked case, the root and its arguments
are optionally embedded under a head which relates the event to its agent.
This head is morphologically encoded in Semitic by the simple binyan:
binyan qal. Two dimensions of markedness are introduced into a derivation
by two additional types of syntactic heads: (a) voice-heads, which modify
voice, and are morphologically realized in Semitic as the passive (pu'al
and huf'al in Hebrew) and middle (nif'al and hitpa'el) binyanim, and (b)
agency-heads, which modify agency, and are morphologically realized in
Semitic as the intensive and causative binyanim. The intensive binyanim
(pi'el, pu'al, hitpa'el in Hebrew) classify the event as an action, and
the causative binyanim (hif'il, huf'al) relate the event to its cause.