Aspect Based Sentiment Analysis With Attention Mechanism And Graph Convolutional Networks

With the advancement of information technology and the rapid development of social media and e-commerce industries,the identity of users is shifting from being receivers of information to producers.More and more people are willing to express their attitudes and sentiments online.This has led to a plethora of comments on people,products,and events on social media,e-commerce,and other internet platforms.These massive amounts of textual information,while expressing various sentiments or attitudes of users,are highly likely to have a certain guidance and impact on subsequent associated users,thus possessing enormous commercial and social value.Currently,how to efficiently and accurately extract viewpoint information from comment texts in an automated manner for text sentiment analysis has become an urgent problem in the fields of computer science and management.Sentiment analysis refers to the process of analyzing,processing and extracting subjective text with sentimental color by using natural language processing and text mining technology.Aspect based sentiment analysis is a fine-grained entity oriented sentiment analysis task aimed at analyzing the sentimental polarity of different aspects of text.At present,research in the field of aspect based sentiment analysis is relatively mature.Based on existing research,the main contents of this study include:(1)A dependency tree enhanced attention model is proposed to address the difficulties in modeling syntactic relationships between aspect items and their contexts with attention mechanisms,as well as the shortcomings of dependency tree structures that make it difficult to focus on aspect items.The model proposes a new aspect oriented dependency tree construction method,which helps the model focus on aspect items by pruning and reshaping the original dependency tree,in order to compensate for the shortcomings of the original dependency tree structure that is difficult to focus on aspect items.The model proposes a dual channel aspect sentiment feature extraction method,which utilizes both aspect oriented attention layers and global oriented attention layers to obtain sentiment features based on aspect context and global context from text embedding,in order to solve the difficulties in modeling syntactic relationships between contexts.At the same time,a new contrastive loss was designed and proposed to enhance the learned attention weights using the aspect oriented dependency tree.The experiment shows that the proposed model achieves better sentiment classification performance than models based on attention mechanism or dependency tree,and can fully leverage the advantages of both attention mechanism and dependency tree,solving the problems in modeling the syntactic relationships between aspect items and their contexts with attention mechanism and dependency tree.(2)A graph guided differentiated attention network model is proposed to address the issue of the high dependence on the quality of dependency tree analysis in aspect based sentiment analysis models based on dependency trees and dense attention weight distribution in attention based models.The core operations of this model are the proposed graph guidance mechanism and attention differentiation operation.The graph guidance mechanism uses the graph structure based on the dependency tree to guide the self attention mechanism to actively learn the attention weight similar to the syntax structure,which helps the model to more accurately and efficiently capture the dependency relationship between words,while reducing the dependence of the model on the dependency tree.The attention differentiation operation encourages the variance of the attention weight learned by the model to increase,making the attention weight distribution tend to be discrete,helping the model learn more discrete and targeted attention weights.The experimental results show that the proposed graph guidance mechanism and attention differentiation operation have positive significance in improving the performance of ABSA.Based on visual cases,the graph guidance mechanism and attention differentiation operation have good interpretability,achieving relatively advanced and robust ABSA performance.