From Graphs to Hypergraphs: Enhancing Aspect-Based Sentiment Analysis via Multi-Level Relational Modeling

ACL ARR 2025 May Submission3182 Authors

Abstract: Aspect-Based Sentiment Analysis (ABSA) demands nuanced modeling of complex aspect-sentiment interactions, a challenge amplified by the limited context in short texts. While graph-based methods have shown promise, they often fall short in capturing higher-order, multi-node relationships, leading them to construct multiple graphs that model fine-grained relationships inherent in language. However, such approaches suffer from poor generalization and increased parameter overhead. To overcome these limitations, we introduce HyperABSA, the first hypergraph-based approach to ABSA, which uniquely leverages a novel hypergraph construction method based on hierarchical clustering with a variance-sensitive threshold. This enables dynamic control over relational granularity via a acceleration based elbow criterion. This single hypergraph framework efficiently captures varying granularities of aspect-sentiment dependencies, while reducing parameter overhead, thereby simplifying prior approaches. Extensive experiments conducted on three public datasets (Lap14, Rest14 and MAMS) demonstrate the effectiveness of our proposed method.

Paper Type: Long

Research Area: Sentiment Analysis, Stylistic Analysis, and Argument Mining

Research Area Keywords: applications

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: English

Submission Number: 3182