Active Learning-based Local Graph Matching for Textual Entailment

This paper presents a robust textual entailment system using the principle of just noticeable dif-ference in psychology, which we call a local graph matching-based system with active learn-ing. First, although an early textual entailment task often involved two rather simple sentences of T (“Text”) and H (“Hypothesis”), the recent textual entailment task often involves multiple / complex / compound sentences in T (andH). In this context, the search for corresponding sub-graphs of T andH is one important task with re-gard to the principle of just noticeable difference where irrelevant items will work as noise for the similarity-based graph-matching approach. Our approach concerns to reduce such noise by linguistic preprocessing being conscious of the topic mentioned by H via subject alignment. Second, by definition of textual entailment task, one interesting task involves the understanding of T which will not be detected by the similarity-based method. Third, one bottleneck of textual entailment task is the performance drop by un-knownwords and named-entities. We explore re-ducing the unknown words and named-entities, incorporating meaning in parentheses / rhetori-cal expressions / semantic roles, and additional feature of language model from deep learning. Our result for RTE2 corpus was 80.49 for macro F1 score, 84.95 for precision for the positive en-tailment, and 79.95 for recall for negative entail-ment