Aspect-based Sentiment Analysis Based on Supervised Learning and Multi-Feature Fusion

Aspect-based sentiment analysis focuses on evaluating the sentiment of aspects, making it a more granular sentiment analysis task. Many existing aspect-based sentiment analysis studies rely solely on either the attention mechanism or the dependency tree to build the adjacency matrix, resulting in inaccurate sentiment analysis results, and only using the output from the final layer of the graph convolutional networks (GCN), which cannot effectively utilize the information learned from each layer of the GCN. Therefore, this paper proposes SLMF-GCN, which stands for Supervised Learning and Multi-Feature Fusion Graph Convolutional Networks. The model uses Bert as the coding layer. In the semantic channel, a supervised learning-based attention mechanism is employed, and the constituent tree is utilized to optimize the attention distribution. Meanwhile, multi-head graph attention is employed for learning each word's representation more effectively. In the syntactic channel, the dependency tree is used to construct the adjacency matrix, the graph convolutional network is used for feature extraction, and the original input of the GCN and the output from every layer are fused to capture the comprehensive features of the sentence. SLMF-GCN fuses the learned semantic and syntactic features to comprehensively comprehend the sentence and achieve aspect-based sentiment analysis. SLMF-GCNl outperforms several existing cutting-edge methods on four datasets.