Research On Chinese Shallow Semantic Analysis Based On BiLSTM-CRF

Deep semantic analysis is a key technology to solve human-machine communication,and the progress made at present is not ideal.Shallow semantic analysis provides a new and feasible solution for solving human-machine communication.Semantic role labeling is the most common method to realize shallow semantic analysis.It aims to label all the semantic roles related to the predicate in the sentence,such as implementation,subject,time and location.Semantic role labeling has the characteristics of concise labeling,clear structure,easy presentation,and technical feasibility.At present,there have been many research results on Chinese semantic role tagging methods,but there are still many challenging problems that need to be solved urgently.Based on the existing annotation methods,this thesis conducts optimization research from two aspects.On the one hand,it is the structural optimization of the sequence annotation model based on deep learning;Another aspect is the targeted fusion of multi-level auxiliary features.The main work done in this thesis is as follows:(1)The BiLSTM-CRF network has many parameters and is difficult to train.This thesis proposes to introduce the maximum pooling technology into the BiLSTM-CRF network to improve the performance of the model.The maximum pooling technology can sample and extract the output of the BiLSTM network to optimize the network structure.Through the comparison of multiple sets of experiments,it is proved that this method can effectively improve the performance of the model.(2)This thesis also tries to introduce multi-level linguistic auxiliary features to improve the semantic understanding ability of the model.At the syntactic level,the dependency syntax and short syntax are introduced;at the lexical level,the part-ofspeech feature is introduced.At the same time,it is proposed to use the average pooling technique to sample multiple features.The purpose of this is to improve the training speed of the model and better release the potential of multi-features,so as to effectively optimize the performance of the semantic role annotation model.The effectiveness of this method has been verified by multiple rounds of experiments.(3)Aiming at the problem that BiLSTM network solves the problem of longdistance semantic dependence of sentences,two ways to improve the model structure are proposed.One is the BiLSTM-CRF network structure that integrates the attention mechanism,which uses the Attention mechanism to reflect the degree of relevance of all words in the sequence and strengthen the semantic understanding of the model;the other is to learn from the TextCNN feature extraction method and propose CNNBiLSTM-MaxPool-CRF Fusion model.The model uses convolution kernels of different sizes to capture the local features of sentences,and then stitches different local features into new features through the average pooling technology.The generated new features have global sentence information,which can improve the model's semantic understanding ability and effectively compensate for the long-distance defects of BiLSTM.Experiments prove that the model has strong practicability.