Research On Aspect-based Sentiment Analysis Method For Web Comment Texts

With the proliferation of Internet technology,users share their attitudes towards things on the network,producing a large amount of comments.These user’s comments reflect the attitudes or emotions of some people and can provide a decision basis for marketing strategy development,online shopping,etc.An aspect is an attribute of an evaluation object and the smallest object to which a user’s attitude in a comment text is directed.The task of analyzing the sentiment tendencies expressed by users with respect to an aspect of the evaluated object is called aspect-based sentiment analysis.For the vast numbers of web comment texts,aspect-based sentiment analysis tasks require aspect word identification before predicting the sentiment tendencies of these aspects,which are unrealistic to be done manually alone.Therefore,aspect-based sentiment analysis of web comment texts using deep learning techniques has become a popular research trend.In order to complete the research task of aspect-based sentiment analysis method for web comment texts,this thesis starts the analysis with two subtasks of aspect extraction and aspect-based sentiment classification,and the main work is as follows:(1)In response to the problems that some sentences in the web comment text dataset are short,the syntactic structure is incomplete,and aspect words show long-tailed distribution,a dual-channel enhancement method of syntactic structure and semantic information is proposed to establish the aspect extraction model SSDR.Firstly,the language model is used to build auxiliary sentences and combined with domain lexicon embedding for semantic enhancement.Secondly,multiple convolutional neural networks are used to extract semantic features and the graph convolutional neural network is enhanced to realize syntactic feature extraction based on syntactic dependency tree and multi-head attention mechanism.Finally,the multi-level gating mechanism is used to adaptively calculate the fusion weights of the two features output from the dual channel to complete the aspect extraction task.(2)Syntactic dependency trees have been widely used in aspect-based sentiment classification tasks,but the current feature extraction and interaction methods are limited to a single feature level,which cannot fully utilize the effective information on the syntactic dependency trees.To address this problem,a multi-level feature extraction algorithm based on syntactic dependency trees is proposed to build an aspect-based sentiment classification model SFEM.First,the syntactic dependency tree is improved on the basis of the known aspectual words in the text to build the syntactic graph.A shallow feature representation of the sentence is derived on the syntactic graph using depth-first search.Then,the syntactic graph is divided into multiple subgraphs and modeled separately when using graph convolutional neural network modeling to extract the deep-level feature representation of sentences.In the end,the sentence representations of deep and shallow level features are fused for sentiment classification.This thesis builds models based on the above research and conducts experiments on several publicly available web comment text datasets,and the final experimental results all prove the effectiveness of the model in this thesis.