Research On Emotion Analysis Based On Syntactic Information

Emotion analysis is an important branch of sentiment analysis that has become a research hotspot in natural language processing due to its wide application.With the development of the Internet,an increasing number of people prefer to express their opinions and emotions on social media platforms such as Weibo and Twitter.Therefore,there are a large number of emotional expression texts.However,compared with the English emotion analysis datasets,Chinese emotion analysis datasets are fewer.Mining Weibo and constructing a high-quality emotion analysis dataset are significant for the study of emotion analysis.In the meantime,because of the short lenght of texts and limited information expression on Weibo and Twitter,existing studies mostly consider textual semantic information and ignore the relationship between words,which results in an insufficient understanding of the text by the model.A few studies use syntactic information.Syntactic information combined with the deep learning model is used as an external feature to enhance the representation of structural information.However,existing studies suffer from the error propagation problem when using syntactic information since the parser can not produce perfect syntactic trees.Therefore,the main contents of this thesis are as follows:(1)Construction of Multi-label Emotion Analysis DatasetBecause of the lack of multi-label emotion analysis datasets,this thesis proposes to construct Weibo emotion analysis dataset.On the one hand,the construction system of the Weibo emotion analysis dataset is proposed.Keeping pace with the times,plenty of highquality Weibo data,keywords and sentence patterns are mined.On the other hand,strict double annotation and expert review are adopted to ensure the consistency of results.In addition,to reduce the workload of labelling,a visual automatic labelling system is developed,which significantly improves labelling efficiency.The final result shows that the Weibo emotion analysis dataset has a high consistency.(2)Emotion Analysis Based on Syntax-aware RepresentationRecently,existing studies suffer from the error propagation problem when using syntactic information in emotion analysis since the parser can not produce perfect syntax trees for colloquial texts.To solve this problem,this thesis proposes an emotion analysis method based on syntax-aware representation.This method directly obtains the syntax-aware wordlevel vector and sentence-level vector representations from the encoder of the upstream dependency parser model,which is injected with the baseline model as additional features.This method avoids the direct use of the syntactic trees.The experimental results show that this method can alleviate the problem of error propagation.It can effectively improve the performance of sentiment analysis model.(3)Emotion Analysis Based on Syntactic Forest Graph ModelTo make full use of the dependency parser model decoder information,this thesis propose to use graph convolutional networks(GCN)to encode dependency trees,in which syntactic forest(the probability matrix of all dependency arcs)is treated as the GCN graph structure.It is used as an additional input and injected into the baseline model.The experimental results on Weibo and SemEva12018 datasets show that after the BERT model obtains syntactic information,the performance of the model can be improved.After being fused with the information represented by the syntax-aware word representation and syntactic forest graph model,the performance of the model is much stronger than the baseline.It can be concluded that by using different levels of syntactic information,the model can make more full use of syntactic information and reduce error propagation.In summary,this thesis proposes to construct a high-quality Weibo emotion analysis dataset.and on the basis,researches an emotion analysis method that integrates different syntactic information.This thesis hopes that the progress in this thesis will contribute to the development of emotion analysis and other tasks in natural language processing.