Research On Key Issues Of Chinese Zero Anaphora For Text Understanding

On the one hand,text Understanding needs to correctly grasp the text's discourse constitutional units and their relationships(logical and semantic relations,topic chain relationships,etc.);on the other hand,it is necessary to understand the overall structure of the discourse and the primary and secondary information of the content expressed.From the perspective of text understanding,this paper conducts a series of researches on Chinese zero anaphora resolution from the perspective of discourse.The main work includes the following three aspects:(1)The construction of Chinese zero anaphora corpus from the perspective of discourse.Chinese zero anaphora corpus resources are scarce,and existing resources mainly sort out and label zero elements and zero anaphora from the syntactic level.But,zero anaphora corpus resources serving text understanding have not been reported.From the perspective of serving text understanding,this paper proposes a Chinese zero anaphora representation scheme for the text understanding.According to the role of zero elements in the elementary discourse unit,the zero element is divided into two categories:the main type and the modified type.Then,the zero anaphora relationship is divided into two dimensions.Finally,based on this representation system,a Chinese zero anaphora corpus is constructed for text understanding,which provides the necessary support for the research of Chinese zero anaphora resolution from the discourse perspective of service text understanding.(2)Chinese zero element position detection.At present,most related researches on zero anaphora resolution have been focusing on the sub-task of zero element resolution,and few studies have been performed on the task of zero element position detection.For text understanding,this paper considers that appearance of the modified zero element is closely related to the local syntactic information,while the main zero element plays a central role in the transfer of contextual topic information.Based on this,while focusing on the local syntactic information,this paper presents a neural model to cast zero element detection as a sequence labeling task,which incorporates both theme and rheme into thedetection model.We further analyze the results from the perspective of text understanding.(3)Chinese zero element resolution.In order to improve the performance of zero element resolution and solve the existing problem not considering the relation between zero element and its antecedent,this paper proposes a zero element resolution model based on the MASK mechanism and the siamese network.Elementary discourse units and sentences are used as joint inputs to optain text-level and sentence-level representations.Experiment results show that our approach can significantly improve the performance of zero element resolution.