Learning Event Expressions Via Semantically Equivalent Projection

Event extraction is one of the most important tasks in information extraction.It is a challenging task due to the high complexity and variety of event descriptions.Event can be defined as a set of actions that occur in a particular time or region and involve one or more participants.Previous work on event extraction can be divided into two categories.One is the supervised machine learning methods relying on a large number of labeled data,which suffered from high cost and poor portability issues.Another is the semi-supervised bootstrapping method.However,the quality of event extraction is highly dependent on the initial seeds and the constraints in the iterative process.Our work falls into the second category,aiming to relax the constraints imposed by predefined syntactic rules.Based on the observation that event phrases often act as a strong indicator of the event type,we propose to utilize the event phrases learned by Huang and Riloff's bootstrapping method^[1]as seeds,on top of which we further expand the event phrases by mining from monolingual and bilingual corpus.We conduct an in-depth study on how to leverage semantically equivalent mapping to harvest new event expression phrases.In this thesis,we propose three methods to address the event phrases extraction problem.?1?Learning event expressions via word-embedding-based semantic projectionFirst,we use distributed semantic representations to cluster words with similar meanings.Second,we acquire reconstructed phrases via retrofitting seed phrases with semantically similar segments.To select retrofitted phrases of high quality among all candidates,we filter phrases based on some linguistic heuristics and leverage a large English corpora to rank these reconstructed phrases.Experimental results demonstrate that our proposed approach can effectively learn new topic event phrases sharing similar meanings to original ones and improve the performance of event recognition.?2?Learning event expressions via paraphrase-based bilingual semantic projectionFirst,we obtain paraphrase pairs via mapping words and phrases across two different languages with machine translation techniques.Second,we evaluate the quality of these pairs by an n-gram language model as well as a linguistic signature annotated corpora.Last,we utilize paraphrases to map similar semantic segments to the seed phrases and harvest new event phrases.Experimental results show that our approach gains further improvement by expanding phrase structures and enriching the event phrases at the semantic level.?3?Learning event expressions via bilingual structure projectionInspired by previous work,we utilize the divergence between two different languages as supplementary information and explore the use of cross-linguistic information to learn event phrases and generalize phrase structure.First,we obtain word alignments between Chinese and English using word alignment tools trained on our parallel corpus.Second,the structural information in phrases is extracted via dependency parser.Third,we combine the structural information and word alignments to project phrases across two languages.Finally,we integrate the proposed bilingual structure projection method into our iterative system.Experimental results show that our approach not only learns new event expression phrases,but also generalizes to yield new phrase structures.Our method overcomes the limitation that event expression extraction severely relies on strict syntactic constraints.