Knowledge Graph enhanced Aspect-Based Sentiment Analysis Incorporating External Knowledge

Autumn Teo, Zhaoxia Wang, Haibo Pen, Budhitama Subagdja, Seng-Beng Ho, Boon-Kiat Quek

Abstract: Aspect-Based Sentiment Analysis (ABSA) is a type of sentiment analysis that could identify and extract various aspects or features from text and determining the sentiment associated with each aspect. ABSA has significant real-world applications, such as providing deeper insights about specific strengths and weaknesses of each aspect contained within text data. Despite notable advancements, ABSA still has room for improvement in completeness, accuracy, and performance efficacy. To tackle these challenges, this research introduces an approach to ABSA that leverages knowledge graphs to improve completeness, accuracy, and performance efficacy. Our key novelty is in being able to incorporate enhancements across multiple stages, including utilising knowledge graphs together with dataset processing, and architectural modelling. Additionally, we offer a complementary overview and analysis of various deep learning heuristics and optimization strategies that could further enhance ABSA performance. Our validation results demonstrate the effectiveness of the proposed knowledge graph enhanced ABSA method across multiple benchmark datasets, with notable boosts to model performance. Importantly, in being model-agnostic, our dataset processing approach could potentially enhance the performance of other ABSA methods in the future.