Research On Chinese Event Extraction And Filling Of Missing Event Argument

Event extraction, a research field of Information Extraction, focuses on how to extract event mention of specific type and its arguments. Nowadays, most of the researches are based on English corpus while Chinese event extraction is still at an elementary stage.In this dissertation, we propose a new method and more effective features for Chinese event extraction based on the existing Chinese event extraction system. Otherwise, we find the full information of an event is often distributed in various parts of the document through the analysis of event extrction results. However, the sentence-level event extraction approaches always ignore these arguments which are out of the sentence, so that a large number of event arguments were missing in our experiment. Therefore, we also propose a theory of filling the missing event arguments based on cross-event inference. The study can be concluded as follow.1. According to the nature of Chinese, this dissertation adopts CRF(Conditional Random Fields) model in trigger detection to solve the problem of the inconsistency between Chinese word segmentation and trigger word boundary. Otherwise, it frist uses cross-event inference in the stage of event type recognition, which expands the feature set form sentence-level information to discourse-level one. Experimental results on ACE2005Chinese Corpus show that our two methods can promote both the accuracy of trigger detection and performance of event type recognition. Compared with the state-of-the-art system, the Fl-measure of our approach can be improved by5.5%and2.5%respectively.2. This dissertation explores the CRF-based event argument extraction approach and summarizes all features into five categories:lexical, semantic, dependency, syntactic and relative-position features. By exploring various features and their combination, we find out that semantic role feature play a important role in our features. Experimental results also show that CRF model has better performance and semantic role is a good indicator for event argument extraction. Compared with the state-of-the-art system, Fl-measure of our event argument extraction approach can be improved by5.1%. 3. To evaluate our argument filling approach, we annotated these missing arguments in ACE2005Chinese corpus firlstly. And then this dissertation proposes a machine learning-base method to fill the missing argument based on the statistic and analysis on the annotated corpus. Our method contains two parts:missing argument identification and classification. The first stage decides whether a missing event argument can be filled while second one decides which argument in other event mention in this document can be used to fill the missing event argument. The experimental results show that the Fl-measure of our method reaches72.97and74.68respectively.