Aspect-level Sentiment Analysis Based On Enhanced Word Vectors And Graph Convolutional Neural Networks

With the popularity of social media and e-commerce,people are more and more inclined to make some comments of things on online platforms.The emotional information contained in these evaluations is of great valuable reference for Internet users,enterprises and the government,so it is very meaningful to conduct aspect-level sentiment analysis according to these text.Researchers have proposed many solutions,but there are still some difficulties in the ASC task.Firstly,from the perspective of the lowest representation of word,there is a lack of methods of word’s representation according to the characteristics of ASC tasks actually.Secondly,from the perspective of text’ feature extraction,the existing models lack enough attention to syntactic information,which makes the models unable to have a comprehensive understanding of sentences.For the above problems,this thesis improves on the basis of existing research,and the main research work is as follows:(1)In this thesis,a position-enhanced word vector’s representation is proposed.Firstly,this thesis proposes to use the BERT model to obtain more accurate text’ word vectors.At the same time,we find that the context which has high correlation with the targeting aspect tends to be concentrated near the aspect word.Therefore,it is proposed to add the positionweighted measurement formula to the model,then multiplying the position weight coefficient obtained by the position-weight measurement formula with the corresponding word vector’s representation obtained by the BERT word vector model,so as to obtain the context’s representation of the targeting aspect word.Experiments show that this way of word’s representation is effective for the improvement of ASC tasks.(2)In this thesis,a novel and efficient GRU-GCN model is designed to extract features of text.We finds that although some words are far away from the aspect words or scattered in distribution,they are actually related to the sentiment of the aspect words,and they are syntactically related.In order to make full use of the syntactic information in sentences,this thesis designs to use a GRU-GCN network.Specifically,the features of text’time series are extracted by the GRU network,and then the GCN based on sentences’ dependency tree is used to find words related to the sentiment of the aspect based on syntactic distance.Experiments show that the proposed model has strong ability of extraction of text’ feature.The model achieves good results.This thesis shows that the method of position-enhanced word vectors can obtain the representation of sentence for specific aspect words and making full use of the positional information of aspect words is conducive to improve the effect of aspect-level sentiment classification.At the same time,the graph convolutional neural network based on sentencedependent tree can make full use of the syntactic information of sentences to achieve more accurate extraction of text’ feature and improve the effect of aspect-level sentiment classification.