Computational Linguistics Colloquium

Thursday, 14 June, 16:15
Conference Room, Building C7 4

Machine Learning for Recognizing Textual Entailment

Recognizing Textual Entailment (RTE) is gaining a lot of attention in these last few years. Given a pair of text fragments, the Text T and the Hypothesis H, the goal of a RTE system is to recognize if T entails H or if it doesn't. As it is a classification problem, many proposed models rely on supervised machine learning techniques to induce or exploit entailment recognition rules described in annotated examples. These models have been very competitive in the Pascal RTE challenges.

This talk will shortly introduce the models presented in these last years during the Pascal RTE challenges and will describe our model for this classification task. The key idea of our model is that entailment recognition rules can be only learnt or exploited if the feature space considers the relations between words in T and H. We use the so-called kernel trick. We define our feature space describing the similarity between examples instead of defining features. Our similarity, the cross-pair similarity, takes into account the relations among T and H using placeholders in the syntactic trees of T and H. This allows the learning model to automatically exploit rewrite rules that are in the annotated examples. Results are encouraging as in the 2nd RTE challenge the system was among the best three. Moreover, the performance of our model is comparable with the first machine-learning-based system when this latter is trained only on the RTE annotated examples.

If you would like to meet with the speaker, please contact Marco Pennacchiotti.