Research On Aspect-Based Sentiment Analysis Algorithm Based On Graph Neural Network

With the rapid development of information technology,text data also began to show an explosive growth,and various online platforms are generating huge amounts of text data every moment.Sentiment analysis uses computer technology to mine text for emotional information,which can provide decision support for users.However,when multiple objects with different sentiment polarity appear in a sentence,sentence-level sentiment analysis is not sufficient to meet the task requirements.The aspect-based sentiment analysis is a fine-grained sentiment analysis that can judge the sentiment polarity of different objects in the text separately,providing a more refined sentiment polarity.Therefore,aspect-based sentiment analysis has become one of the popular research directions in the field of natural language processing,and GNN in deep learning have become a hot spot for solving this task.GNN operate on the adjacency matrix generated by the dependency tree,but most of the existing models ignore the noise of the dependency tree and do not fully consider the importance of aspect objects,and the accuracy needs further improvement.Therefore,this paper assembles GNN,BERT models and attention mechanisms to improve the above problems:In the first part,an dual-channel and multi-granularity gated graph attention network for aspect-based sentiment analysis(DMGGAT)is proposed.Previous graph neural networks use dependency tree to extract syntactic information,but the quality of dependency trees generated by different parsing dependents varies,so it may cause noise to the model.We combine the BERT model and the GAT network to form a two-channel framework for learning semantic and syntactic features.the BERT model not only supplements the semantic information lost by the GAT computation,but also provides syntactic information to weaken the noise of the dependency tree.In addition,we design an attention mechanism based on aspect objects to extract contextual sentiment information associated with aspect words and generate coarse-grained representations by averaging pooling.Finally,a gating module is used to optimize the fusion of feature information from both channels.The experimental results demonstrate the validity of the proposed method and the effectiveness of each module by ablation experiments.In the second part,an aspect-guided multi-graph convolutional network for aspectbased sentiment analysis(AGGCN)is proposed.In the first part,the dependency graphs all have values of 1 or 0.However,in text data,usually the closer the word to the aspect word the greater the relevance.For this purpose,this section manually modifies the value of the dependency graph by basing on the distance of the aspect words from other words,so that the words closer to the aspect words have more weight.As in the first part,the noise of the dependency tree needs to be weakened.Previous studies utilize the score matrix of multiheaded self-attention as a semantic graph.In this section,we propose to add aspect information to the multi-headed self-attention calculation to obtain an aspect-based score matrix.Finally,the features of the three graphs are captured by GCN.The experimental results on three datasets show that AGGCN achieves advanced results.