Research On Deep Learning Based Chinese Pronoun Resolution And Its Application In Question Answering

The phenomenon of anaphora and ellipsis bring challenges to machines understanding natural language, therefore pronoun resolution, which can solve the problem of semantic vacancy, has been paid more and more attention to. This thesis proposes novel methods to resolve Chinese pronoun with deep semantic features represented by word embedding and achieves a completing-semantic module in question answering system.Firstly, this thesis focuses on Chinese overt pronoun resolution and completes the mention pair model. In this part, the thesis makes an improvement on explicit semantic features extracted from experience, and combines with hidden semantic features represented by word embedding as the final features of the mention pair. Evaluated on Onto Notes5.0 corpus, the system employing combinational features performs better than employing explicit or hidden features separately.Secondly, this thesis employs deep-learning approaches to identify and resolve Chinese zero pronoun. The thesis proposes a novel identification algorithm based on bi-direction recurrent neural network without syntax parsing, meanwhile, achieves zero pronoun resolution with mention pair mo del of overt pronoun resolution by ignoring explicit features and supplementing embedding features of verbs and object, then proposes a novel architecture with long-shortterm memory network to capture deeper semantic information of mention pairs. The results show that, compared with baseline system, the methods this thesis proposes have superiority in identification and resolution subtask.Finally, this thesis implements a question answering system that recovers the important omitted semantics of current question by Chinese pronoun resolution algorithm and co-occurrence language model from short-term history questions before retrieving. The instances explain that completing-semantic module can improve the performance of question answering system.