Research On Chinese Event Coreference Resolution

The main task of event coreference resolution is to find events coreference chains, which is widely used in many Natural Language Processing(NLP) applications, such as event extraction, topic detection, question answering and text summarization, etc. Currently, there are few and fragmented studies on Chinese corpus.This paper focuses on the following aspects: event coreference resolution, global optimization of event coreference resolution and joint inference of event coreference resolution. The main contents are as follows:(1) Event coreference resolution based on feature methodIn pairwise model, every two events are made up into event pairs. According to Chinese characteristics, such as flexible expression and word polysemy, etc., many useful features are proposed including phrases, sentences and distances, etc., to construct a machine learning-based pairwise model to resolve coreference events.(2) Global optimization for event coreference resolutionThe pairwise model ignores the internal relations between events, causing the inconsistency chains of event coreference easily. This dissertation proposes the global inference method to reduce the contradictions of event coreference chains which are caused by the classifier. This method contains several novel and effective constraints, such as argument role, event distance and trigger semantic, etc., further improving the performance of event coreference resolution.(3) Joint inference methods for event coreference resolutionPresent researches mainly consider the event coreference relationships, and the results are contradictory in other fields. This dissertation proposes joint inference methods including two parts: joint inference for event coreference resolution and temporal relation recognition, joint inference for event coreference resolution and argument recognition. The former focuses on the event relations. With the semantic and discourse information provided by event temporal relations, the performances of event inference resolution increase obviously with high consistency on event temporal relations. The latter joints the event argument recognition and event coreference resolution, which can reduce the cascade errors in pipeline model. Experimental results show this method is helpful for the two tasks.This dissertation focuses on Chinese corpus to resolve coreference events and proposes three approaches to improve its performance. Our work will be beneficial for further research work in this field and related fields.