Yonatan Zarecki, M.Sc. Thesis Seminar
Tuesday, 17.4.2018, 12:30
Labeling textual data by humans can become very time-consuming and expensive, yet critical for the success of an automatic text classification system.
In order to minimize the human-labeling efforts, we propose a novel active learning (AL) solution, which does not rely on existing sources of unlabeled data.
It uses a small amount of of labeled data as the core set for the synthesis of useful membership queries (MQs) - unlabeled instances synthesized by a computer, for human classification.
Our solution uses modification operators, functions from the instance space to the instance space that changes the input to some extent.
We apply the operators on the core set, thus creating a set of new membership queries.
Using this framework of modification operators we look at the instance space as a search space and apply search algorithms in order to create desirable MQs.
We implement this framework, along with its modification operators, on the textual domain and test it on text classification tasks.
We show an increase in classifier performance as more MQs are labeled and incorporated in the training set in comparison to other options of gathering labeled examples in several datasets.
In our experiments we even show that using our MQs can be competitive to using a more traditional pool-based active-learning approach, which requires an additional pool of unlabeled instances.
To the best of our knowledge, this is the first work on membership queries in the textual domain.