Aspect-based Sentiment Analysis Model Based on Multi-hop Information

Textual aspect-level sentiment analysis aims to predict the sentiment tendencies of specific aspects within sentences. In recent years, the majority of methods have primarily focused on the semantic information within sentences, employing traditional attention mechanisms in combination with recurrent neural networks to capture contextual information, while not adequately considering the importance of relevant syntactic constraints. In this paper, a combination of dual graph convolutional networks is employed as the foundation for extracting semantically and syntactically related features. With an increase in the number of network layers and iterations, the features of all nodes within the same connected component tend to converge gradually after multiple convolution operations. To tackle the issue of over-smoothing, this paper introduces a multi-hop information propagation mechanism that extends the receptive field of each layer within the graph convolutional network. Additionally, a novel method for encoding syntactic information is presented, based on the structure of dependency trees. This method utilizes a multi-head attention mechanism to capture the interaction between semantics and syntax. Experimental validation was conducted on three publicly available datasets (Restaurant, Laptop, and Twitter), which include textual data with varying sentiment tendencies (negative, neutral, and positive). The experimental results underscore the effectiveness of the proposed approach.