Research On Sentiment Expression For Aspect-Based Sentiment Classification

With the vigorous development of the Internet industry,consumers are no longer satisfied with the traditional forms of offline shopping.E-commerce shopping platforms such as Taobao and jd.com have become an essential new choice.When consumers experience online shopping,they also produce a large number of evaluation data on ecommerce services.These evaluation data are not only helpful to improve sellers’understanding of their own goods and services,but also become the important reference for other consumers.Different from the traditional sentence level text sentiment classification,aspect-based sentiment classification(ABSA)focus to the specific attributes in sentences.For complex e-commerce comment texts,the existing ABSA methods are still insufficient.Existing studies often rely on the relationship between aspect words and context.However,for such complex comments,it is easy to have a lot of additional noise.Therefore,this dissertation proposes research on sentiment expression for aspect-based sentiment classification.For a given aspect word and the comment text,obtain the sentiment simplified text related to the attribute word,and assist the comment text for sentiment analysis.Starting from the acquisition method of sentiment expression,this study is divided into three levels:1)Research on boundary detection for ABSA:acquiring sentiment expression text by text extraction;2)Research on sentiment expression generation for ABSA:obtain sentiment expression text by text generation;3)Research on sentiment expression and dependency information for ABSA:considering the syntactic and semantic information of the text,obtaining the sentiment expression text.The specific research work is as following:First,aiming at the problem that the sentiment expression of attribute words accounts for only a small part of complex comments and is easily affected by external noise,this dissertation proposes an attribute level sentiment classification method based on boundary detection.First,collect and process Chinese e-commerce comments from three different fields of Taobao e-commerce and English comments on SemEval’s restaurants in 2014,2015 and 2016 task,and then mark the sentiment expression position of each attribute of each comment.Second,establish a text extraction model based on sequence tagging.The attribute words and comment text are taken as input.The words belonging to sentimental expression in the comment are extracted by sequence tagging and used as the simplified text of sentimental expression.Finally,combine the extracted simplified text of sentimental expression with the original comment for sentimental classification.Experiments show that the method proposed can effectively alleviate the influence of other noise in complex comments,make full use of the sentiment expression information related to attribute words,and improve the effect of attribute level sentiment classification.Second,aiming at the problem that the related sentiment expression clauses of attribute words in comments are limited by the speech expression of the original text,this dissertation proposes an attribute level sentiment classification method based on sentiment expression generation.First,attribute words and comment text are used as input to learn text features through attention mechanism.Second,construct the text generation module.In the text generation training,the special mask mechanism in unilm is used to construct the text generation model based on Bert.At the same time,the implicit variable support of attribute words and sentimental words is added to the text generation module to strengthen the sensitivity of the generated text to attribute words and maintain the sentimental consistency between the generated simplified text and attribute words in the original comment text.Experiments show that the text generation method can effectively alleviate the limitations of text extraction,simplify the sentimental expression,and make the text easier for machine understanding.The attribute level sentiment classification method based on sentiment expression proposed in this dissertation is generally higher than the baseline experiment in the accuracy of attribute level sentiment classification.Finally,according to the problem that sentiment expression is always too complex to be simplified,this dissertation proposes an attribute level sentiment classification method based on sentiment expression and dependency.First,a natural language processing tool is used to analyze the comment text to obtain the syntactic dependency of the comment text.Second,the related dependencies are extracted and serialized centered on attribute words.Then,attribute words,comment text and dependencies are taken as inputs.In the text generation model based on Bart,the sequence embedding vector of dependencies is added to strengthen the semantic understanding of attribute words in comment text in the process of text generation and learning.Meanwhile,the implicit modules of attribute words and sentimental words are added to strengthen the attention to attribute words and the consistency of sentimental tendency expression,respectively.Experiments show that the attribute level sentiment classification method combined with dependency and text generation proposed in this dissertation can effectively analyze the syntactic structure of comment text,and use the contextual semantic information of attribute words to improve the generation quality of sentiment expression and simplify text.Compared with the baseline experiment,this method has higher accuracy of attribute level sentiment classification.