CGT: A Clause Graph Transformer Structure for aspect-based sentiment analysis

Abstract

In the realm of natural language processing (NLP), aspect-based sentiment analysis plays a pivotal role. Recently, there has been a growing emphasis on techniques leveraging Graph Convolutional Neural Network (GCN). However, there are several challenges associated with current approaches: (1) Due to the inherent transitivity of CGN, training inevitably entails the acquisition of irrelevant semantic information. (2) Existing methodologies heavily depend on the dependency tree, neglecting to consider the contextual structure of the sentence. (3) Another limitation of the majority of methods is their failure to account for the interactions occurring between different aspects. In this study, we propose a Clause Graph Transformer Structure (CGT) to alleviate these limitations. Specifically, CGT comprises three modules. The preprocessing module extracts aspect clauses from each sentence by bi-directionally traversing the constituent tree, reducing reliance on syntax trees and extracting semantic information from the perspective of clauses. Additionally, we assert that a word’s vector direction signifies its underlying attitude in the semantic space, a feature often overlooked in recent research. Without the necessity for additional parameters, we introduce the Clause Attention encoder (CA-encoder) to the clause module to effectively capture the directed cross-correlation coefficient between the clause and the target aspect. To enhance the representation of the target component, we propose capturing the connections between various aspects. In the inter-aspect module, we intricately design a Balanced Attention encoder (BA-encoder) that forms an aspect sequence by navigating the extracted phrase tree. To effectively capture the emotion of implicit components, we introduce a Top-K Attention Graph Convolutional Network (KA-GCN). Our proposed method has showcased state-of-the-art (SOTA) performance through experiments conducted on four widely used datasets. Furthermore, our model demonstrates a significant improvement in the robustness of datasets subjected to disturbances.