Event Entailment Extraction Based On Clustering Method

The concept of textual entailment originated in logic. It is also an important issue in linguistics, in particular, semantic research area. In recent years, with the advances in artificial intelligence research, especially in the field of natural language processing, textual entailment become a popular topic in computer science. With the mechanism of textual entailment, the performance of various natural language processing tasks will be improved, including information extraction, information retrieval, document summarization, question answering and text-to-scene conversion system. Therefore, the study of textual entailment is very important in natural language processing.Event entailment is part of textual entailment which focuses on verbs. Methods to extract event entailment relations in the past are not only lead to normal accuracy but also high time complexity. First we use clustering method to classify all sentences extracted from corpus into different classes and then try to find entailment relations in such classes, which will significantly decrease time cost. At the same time we also abandon the traditional, statistical KL-like distance measure or metric of mutual information, but use a resource-based semantic similarity measure to score for the verb pairs, not only further reduce the time overhead, also achieved good accuracy.This paper also shows the sample of the event entailment relation library produced by our new method in both Chinese and English. Several reviewers are invited to evaluate samples randomly selected from the library. The result showed that the library contains not only many common entailment relations but also some relations which not exist in artificial library and the accuracy of our library achieved the average level of automatically generated entailment relation library.Finally, this paper describes an entailment search system which allows users to query verbs they want to find entailment relations about them. Its back-end database is constructed by the results of our experiments. The system distinguishes normal users and registered users. Normal users can only query and see the result while registered users have certain rights, including remove incorrect result from the database and correct inappropriate usage examples of verbs. The search system greatly enhanced the availability of the generated library. If registered users continue evaluating and modifying the database, the database will eventually reach a very high accuracy.