Research On Chinese Event Inference Approach Based On Markov Logic Networks

Under the environment of the explosive growth of data, information extraction has become an important method to obtain valuable information from vast amounts of resources. Event extraction is an subtask of information extraction, which aims to present the unstructured text with event information in a structured form. It plays an important role in many Natural Language Processing (NLP) applications, such as question answering, automatic summarization and information retrieval, etc. One disadvantage of those existing event extraction approaches is that they mostly regard event mention as isolated individuals and then ignore inner relationship among event mentions.This dissertation focuses on the following three aspects:Chinese event trigger inference, Chinese event argument inference and joint Chinese event inference. The main contents are as follows:Firstly, due to the ellipsis problem in Chinese, this dissertation proposes a Chinese event trigger inference approach based on Markov Logic Networks (MLN). This approach employs discourse consistency theory, morphological structures and compositional semantics to infer triggers. Experimental results on the ACE2005 Chinese Corpus show that compared to the baseline, the F1-measure of our approach can be improved by 3.65% and 2.51% in trigger identification and event type determination, respectively.Secondly, current Chinese argument extraction approaches mainly use syntactic structure as the major feature to describe the relationship between the trigger and its arguments. However, they suffer much from those arguments which are far from their triggers. To address this issue, this dissertation brings forward a novel argument inference mechanism based on MLN, and employs discourse consistency theory and information of entity semantics to infer arguments. Experimental results on two event themes (conflicting theme and justice theme) show that compared with the baseline, our method improves the F1-measure by 6.84% and 5.7% in argument identification and role determination, respectively.Finally, traditional event extraction systems adopt pipeline architecture, which causes cascading errors. To address this issue, this dissertation proposes a joint Chinese event inference model based on MLN. This model employs discourse level information, adding coreference consistency and arguments quantity inference rules, which exploits useful information in argument extraction in turn to help trigger extraction, and then partial missing triggers are recovered. Experimental results show that compared with the baseline, our model improves the F1-measure by 1.68%,1.63%,1.92% and 1.89% in trigger identification, event type determination, argument identification and role determination, respectively.This dissertation brings forward the Chinese event inference approaches based on MLN and experimental results testify the effectiveness of our approaches, which will be beneficial for further research work in this field and related field.