Knowledge Graph-Guided and Dependency Distance Enhanced Multi-Graph Networks for Aspect-Based Sentiment Analysis

Abstract

Aspect-based sentiment analysis (ABSA) determines the affective polarity of given aspects within a sentence by recognizing their affective tendencies. Many recent approaches employ graph convolutional networks to extract information from dependency trees to solve ABSA tasks. However, it remains a worthwhile problem to better utilize semantic and syntactic key information contained in dependency trees for improving ABSA. In this paper, we propose the Knowledge Graph-guided and Dependency Distance Enhanced Multi-Graph Network (KDE-MGN) to face this challenge. Specifically, We designed a multi-graph generation module to strengthen the acquisition of syntactic and semantic information. The module constructs a Dependency Distance Weight Matrix (DWM) and Dependency Distance Mask Matrix (DMM) based on dependency distances between words. We combine DWM with Dependency Neighborhood Matrix to construct a Distance Enhanced Syntactic Graph (DESyG) for more comprehensive syntactic information. Moreover, to obtain more semantic information related to aspects, we introduce an Aspect Enhanced(AE) module to compute attention scores matrices for sentences. We then combine the DMM with an attention matrix to construct a Distance Enhanced Semantic Graph (DESeG) as a way to learn more comprehensive semantic information both locally and globally. Meanwhile, we also present a knowledge graph enhancement module to guide KDE-MGN to capture more important sentiment information in context. Extensive experiments on four baseline datasets show that our proposed model has state-of-the-art performance.

Keywords: Aspect-based sentiment analysis, Graph neural network, Dependency distance, Knowledge graph