Integrating Contextual Target Attributes for Target Sentiment Classification

In this article, we delve into the realm of sentiment analysis, specifically focusing on the role of syntactic information in capturing target-related semantics. We explore how graph neural networks (GNNs) can be applied to encode the syntax graph of a review context, facilitating the capture of beneficial clues for sentiment analysis. Our proposed model, Heterogeneous Graph Convolutional Networks (HIG2CN), integrates attention mechanisms and relative position weighting layers to effectively enhance context understanding. By leveraging the interactions among background information, property information, syntactic information, and contextual word correlative information, HIG2CN can adaptively control the information flow during the aggregation process.
To begin with, we highlight the significance of syntactic information in improving sentiment analysis performance. We observe that if the semantics graph is not considered, the performance drops sharply, as the semantics graph includes correlation scores between every two words, which can capture more comprehensive contextual information. However, attention mechanisms often struggle to fully capture target-related semantics contained in some context words. To address this challenge, we propose incorporating GNNs into the model to encode the syntax graph of the review context.
By applying a multi-layer graph convolutional network (GCN) on the syntax graph, HIG2CN can effectively capture the dependencies between the target and crucial words in the context. The relative position weighting layer calculates the weight of each context word regarding its relative distance to the target words, emphasizing the importance of proximity in determining sentiment. Moreover, the heterogeneous convolution layer combines information from different types of nodes in the graph, enabling the model to capture both local and global contextual relationships.
In summary, HIG2CN represents a promising approach to sentiment analysis by leveraging syntactic information and GNNs to enhance context understanding. By integrating attention mechanisms and relative position weighting layers, our proposed model can effectively capture target-related semantics in review contexts, leading to improved performance in sentiment prediction tasks.