Research On Aspect-level Sentiment Analysis Technology Based On User Review

User reviews texts in social networks are rich in emotional information,and users’ comment objects often contain one or more different aspects.A comprehensive and accurate extraction of users’ emotional polarity on different aspects can help the government to monitor public opinion in a targeted manner and make accurate responses.Aspect sentiment analysis technology aims to mine emotional information that users express towards different aspects.This technology involves multiple subtasks that affect the development process,including aspect term extraction and aspect category sentiment Analysis.In this paper we proposed corresponding improvement methods for these two tasks.The main work is as follows:(1)An aspect term extraction method based on constituency parsing and Attention-BiLSTM is proposed in this paper.To address the performance deficiency of most existing aspect term extraction models in phrase-level aspect term extraction,firstly the constituent parsing technique is used to generate a constituency parse tree to capture the important syntactic structures in the sentence,and then,to help the task of target term boundary identification,a lattice structure is used to construct a constituency lattice from the constituency parse tree,and the Attention-BiLSTM model based on the constituency lattice encoding is proposed to generate more efficient vector representation,finally the CRF model is used to achieve the extraction of aspect terms.The performance of the method is evaluated on two datasets,and the experiment results show that the method improves 3.88%6.93%and 0.11%-4.69%on the two datasets compared with various baseline models,and is able to extract aspect terms in user comment texts more accurately.(2)An aspect category sentiment analysis method based on pre-trained BiLSTM and syntax-aware Graph Attention Networks is proposed in this paper.In response to the current situation of relative scarcity of aspect-level sentiment analysis annotated data,a transfer learning approach is proposed.firstly,use a BiLSTM model to pre-train on a document-level sentiment analysis dataset,and transfer the obtained pre-training parameters to a Bi-LSTM model for the aspectlevel task,to achieve effective learning of relevant domain knowledge from a large amount of document-level sentiment data.Then a syntax-aware Graph Attention Network is proposed to Implement an aspect-level sentiment analysis task by using constituent parsing technique and Graph Attention Networks,making full use of the constituent structure and semantic information of the text,and combining the prelearned knowledge from the pre-trained Bi-LSTM model.The performance of the method is evaluated on five user reviews texts datasets,and the ablation experiments demonstrate that the method performs best compared to all baseline models and can perform the aspect category sentiment analysis task more accurately.In this paper,the effectiveness and accuracy of the proposed method is demonstrated through detailed experiments using a variety of publicly available datasets of user review texts.The experiment results show that the aspect term extraction method and aspect category sentiment analysis method proposed in this paper obtain better results compared with multiple comparison models.