Research On Target-oriented Opinion Words Extraction Based On Deep Learning

Aspect level sentiment analysis is the research focus of text emotion analysis task.Its basic sub tasks include target word extraction,opinion word extraction and aspect level emotion classification.The traditional target word extraction and opinion word extraction are completed independently,which separates the relationship between target words and opinion words.To solve this problem,the latest research work proposes a target-oriented opinion word extraction task.Specifically,the target oriented opinion word extraction task refers to extracting the corresponding opinion words according to the specific target words.When there are multiple target words in a sentence,it is still a very challenging task to accurately capture the relationship between the target words and the corresponding opinion words.This paper studies the target-oriented opinion word extraction task.The main research contents and innovative work are as follows:Firstly,target-oriented opinion word extraction models mostly focus on using semantic features to complete the extraction task,ignoring the syntactic features of sentences.To solve this problem,this paper combines graph convolution network(GCN)and graph attention network(GAT)to improve the neural network model(Target-oriented Opinion Words Extraction with Target-fused Neural Sequence Labeling,IOG)integrating the semantic features of target words,and puts forward the IOG-GCN-GAT model.Firstly,the context semantic features specific to the target word are learned by using IOG;Then,GCN and GAT are used to fully mine the syntactic features on the syntactic dependency tree;Furthermore,the semantic features and syntactic features are spliced to obtain the final context coding representation;Finally,softmax is used to complete the target oriented opinion word extraction task.By adding syntactic supervision signals,the model effectively improves the effect of target oriented opinion word extraction..Secondly,the syntactic dependency tree used by IOG-GCN-GAT is obtained by parsing the sentence.Different target words in the sentence follow the same syntactic tree,so it is unable to make full use of the characteristics of the target words.To solve the above problems,this paper integrates bidirectional long-Short term memorry(Bi LSTM)and bidirectional graph convolution neural network(Bi GCN)on aspect-oriented dependency tree(ADT),and proposes Bi LSTM-ADTBi GCN model.For each specific target word,the model constructs a structure tree centered on the target word according to the dependent syntax tree of the sentence.After using Bi LSTM to learn contextual semantic features,the model uses Bi GCN mine the syntactic features on ADT.Bi GCN can spread the target word information to the context from top to bottom,and aggregate the context information to the target word from bottom to top.The experimental results show that the target-oriented word extraction effect of Bi LSTM-ADTBi GCN can be improved.Lastly,the syntactic information introduced in the above model is very dependent on the syntactic dependency tree parsed by the syntactic parser,and the parsing result of the syntactic parser is not the most efficient.To solve the above problems,based on the IOG model,this paper integrates a graph convolution neural network modeled on the FT-Roberta induced tree(FRIT)of Roberta pre-training model,and proposes the IOG-FRITGCN model.Specifically,after fine tuning Roberta according to the target-oriented opinion word extraction task,the FRIT that is more in line with the task requirements is probed by using the perturbed masking strategy.Compared with ordinary syntactic dependency tree,FRIT can shorten the distance between target words and opinion words more effectively.The experimental results show that IOG-FRITGCN can effectively improve the effect of targetoriented opinion word extraction.