Local Dependency-Enhanced Graph Convolutional Network for Aspect-Based Sentiment Analysis

Abstract

The task of aspect-based sentiment analysis (ABSA) is to detect the sentiment polarity toward given aspects. Contemporary methods predominantly utilize graph neural networks and incorporate attention mechanisms to dynamically connect aspect terms with their surrounding contexts, resulting in more informative feature representations. However, these methods only consider whether there are dependencies between words when introducing dependencies, ignoring that dependencies between different sentiment words have different effects. Neglecting this could introduce noise and negatively impact the model’s performance. To overcome this limitation, we introduce a novel approach called the local dependency-enhanced graph convolutional network (LDEGCN). Our method combines semantic information and dependency relationships to better capture the affective relationships between words. Specifically, we integrate sentiment knowledge from SenticNet to enrich the sentence’s dependency graph and thoroughly explore the dependency types between contexts and aspects to focus on particular dependency types. The local context weight (LCW) method is employed on the dependency-enhanced graph to emphasize the importance of local contexts, thereby mitigating the issue of long-distance dependencies. Through extensive evaluations of five public datasets, the LDEGCN model demonstrates significant improvements over mainstream models.

Keywords

• aspect-based sentiment analysis
• graph convolutional network
• multi-head attention
• affective knowledge
• dependency types