Research And Implementation Of Chinese Zero Pronoun Resolution Method

Zero Pronoun is a special linguistic phenomenon,which refers to a linguistic unit that is omitted to ensure the continuity of the language and can be inferred from the context.The omitted language unit bears the corresponding syntactic component in the sentence,and may form a referential relationship with the language unit such as one or more noun phrases in the preceding text.Because of the language characteristics of Chinese,zero pronoun is very common in Chinese.Therefore,Chinese zero pronoun resolution is very important for Chinese natural language understanding.It is one of the important basic tasks of Chinese natural language processing.It is also very important for other downstream tasks such as syntax analysis,semantic role labeling,machine reading comprehension,and machine translation.Zero pronoun resolution in Chinese can usually be divided into two sub-tasks,one is the recognition of zero pronouns,and the other is the resolution of zero pronouns.This thesis uses the neural network method to research on these two sub-tasks separately,and finally uses a joint model method to combine them.The main work of this thesis includes the following three aspects:(1)A Chinese Zero Pronoun Recognition Model Based on Deep Learning.This thesis proposes a zero pronoun recognition method based on deep neural networks,which uses the attention mechanism to capture semantic information in the context,assigns higher weight to words that contain more semantic information in the sentence,and uses Tree-LSTM to mine Syntactic structure information.Finally,the zero pronouns are identified by the fusion features of the two.The experimental results show that the method proposed in this thesis can effectively improve the recognition effect of Chinese zero pronouns.(2)Chinese Zero Pronoun Resolution Model Based on Bert and Biaffine attention mechanism.In this thesis,the Bert is used to replace the traditional LSTM feature coding layer to obtain more accurate and rich feature representations.When representing candidate antecedents,the self-attention mechanism is used to model the global information of candidate antecedents.Finally,the classification is performed by the Biaffine attention mechanism.The experimental results show that the Bert can fully extract the feature information in the text,and the Biaffine attention mechanism can better learn the interaction information between the zero pronouns and the candidate antecedents forclassification.The model proposed in this thesis effectively improves the resolution.(3)End-to-end joint Chinese zero pronoun resolution model based on Bert.Based on the previous work of identifying and resolving Chinese zero pronouns,this thesis first implements a pipeline Chinese zero pronoun resolution system.It is considered that this pipelined approach will cause cascading error,and ignore the correlation in the process of identifying and resolving zero pronouns.In this thesis,Bert is used as a feature extraction layer,and joint learning is used for Chinese zero-pronoun resolution.The experimental results show that the end-to-end joint model proposed in this thesis effectively improves the shortcomings of the pipeline model,and improves the Chinese zero-pronoun resolution.