Research Of End-to-End Aspect Based Sentiment Analysis Based On Syntactic Structure

In the era of the Internet,online service platforms have become the main channel for obtaining information and expressing opinions.The text data generated therein contains rich sentiment information,which is important to governments,enterprises and individuals.The aspect-based sentiment analysis can capture sentiment information corresponding to different aspects in a sentence,while the end-to-end aspect-based sentiment analysis aims to simultaneously obtain the aspect and its corresponding opinions,thus has important research significance and practical value.The text sentiment feature extraction is the basis of aspect-based sentiment analysis.Firstly,this thesis analyzes how to extract text sentiment feature with the help of syntactic information.The existing syntactic-based methods have two major defects,one is the inaccuracy and redundancy of the syntactic information will lead to error propagation and noise;another is the long syntactic distance between words will make it difficult to construct a connection between the aspect and the corresponding opinion.This thesis proposes a Relational Sequence Encoding Graph Attention Network(RSE-GAT)to extract the sentiment feature.By constructing aspect-specific dependency tree,the syntactic distance between aspect and context is shortened,and by generating aspect-specific dependency representations in the Relational Sequence Encoding layer,the effects of noise and error propagation are mitigated.In order to enhance the ability of sentiment feature extraction,the node representation composition layer is used to interact the word representation and the dependency relation representation.The end-to-end Aspect Sentiment Quad Prediction task needs not only to extract high-quality text sentiment features,but also to correctly output four kinds of sentiment element into a quadruple,which mainly faces the following two challenges.The first one is how to deal with the implicit aspect and opinion in the sentence;the second one is sow to use the semantic information of aspect category to enhance the model.This thesis proposes a syntactic-based Aspect Sentiment Quad Prediction model RSE-GAT-ASQP,which propose a new modeling paradigm to adapt to implicit aspect and opinion.This model use "screening-confirmation strategy" to improve the representation effect of aspect category’s semantic information and the processing ability of the model in implicit scenes.Furthermore,by using sentiment polarity encoding to provide auxiliary information,the model performance of extracting aspect term and opinion term by sequence labeling is improved.Experiments on aspect sentiment classification tasks show that RSE-GAT can extract text sentiment feature better than 16 representative baseline methods.RSE-GAT-ASQP achieved better results than the four baseline models in two public datasets of the ASQP task.In addition,the ablation experiment further verifies the validity of the innovation components of the text sentiment feature extraction method and the aspect sentiment quadruple prediction model.