Aspect-Level Sentiment Analysis Based on Part-of-Speech Enhanced and Graph Convolution Network

Although graph neural networks based on syntactic information have achieved good results in aspect-level sentiment analysis, these methods still have certain limitations and often struggle with accurately processing the semantic and syntactic information between aspect words and opinion words. Therefore, we propose an aspect-level sentiment analysis Based on part-of-speech enhanced and graph convolution network(PSEGCN). Specifically, we design a part-of-speech attention mechanism combined with self-attention to obtain an enhanced semantic adjacency matrix. By focusing on important part-of-speech words, this approach strengthens the semantic correlation between aspect words and opinion words while reducing noise from irrelevant words. Next, we construct an aspect-distance-based adjacency matrix from the original dependency tree and combine it with the dependency relationship adjacency matrix. This fused matrix allows for a more comprehensive learning of the syntactic information of aspect words. Additionally, we use two GCN networks to encode the information from the two matrices. Finally, we use a gated mechanism function to better integrate syntactic and semantic information for sentiment classification. Our model achieved accuracy rates of 0.7816 on the laptop dataset and 0.7565 on the Twitter dataset, outperforming all baseline models, which demonstrates the effectiveness of our model.