Research On Semi-supervised Chinese Event Extraction

Event extraction, a research field of Information Extraction, focuses on how to extractevent mention of specific type and its arguments. Nowadays, most researches concernsupervised models and only a few focuses on English semi-supervised or unsupervisedevent extraction while semi-supervised Chinese event extraction is still at an elementarystage.This dissertation focuses on three aspects of semi-supervised Chinese event extraction,i.e. extraction model, event pattern and event inference. The study can be concluded asfollow.At first, due to the differences between English and Chinese language, it proposes atwo-view-based bootstrapping approach to extract events. According to a small set of seeds,it applies a cross filtering method to two views, which based on document relevance andsemantic similarity, and extracts new patterns in each iteration. Experimental results onACE2005Chinese Corpus show that the F1-measure of our approach can be improved by8.2%and3.6%respectively, compared with the document relevance view and semanticsimilarity view.Then, due to the issues of polysemy of triggers, it puts forward a semantics-basedpattern optimization method. First, an argument-based pattern filtering approach improvesthe accuracy of the pattern. Second, a pattern conversion approach based on syntacticstructure enhances the applicability of the pattern. In addition, a mixed pattern approachfurther clarifies the semantics of triggers. The experimental results show that theF1-measure of our system can be improved5.3%.At last, the performance of semi-supervised Chinese event extraction depends on thequality of seed patterns. However, the coverage of those seed patterns is limited and lots ofsparse patterns cannot be identified. To solve this issue, it proposes an event inferencemechanism based on co-reference events and relevant events, following the theory of discourse consistency in a topic. According to event mentions extracted this mechanisminfers others which have the co-reference or relevance relations in the same document. Theexperimental results show that the performance of our system can be improved11.1%.This paper proposes optimization methods for semi-supervised Chinese eventextraction system. The experimental results prove the validity of the methods. They areconducive to the development of semi-supervised Chinese event extraction technology.