CMV-Fuse: Cross Modal-View Fusion of AMR, Syntax, and Knowledge Representations for Aspect Based Sentiment Analysis

ACL ARR 2026 January Submission6483 Authors

Keywords: Aspect Based Sentiment Analysis, Multi-View Fusion, Contrastive Loss, Abstract Meaning Representations, Dependency Trees, parsing and related tasks, dependency parsing

Abstract: Natural language understanding benefits from integrating complementary linguistic perspectives spanning syntax, semantics, and external knowledge. However, most existing Aspect-Based Sentiment Analysis (ABSA) models rely on isolated linguistic views or ad hoc fusion strategies, limiting their ability to jointly reason over diverse structural representations. We propose CMV-Fuse, a Cross-Modal View fusion framework that systematically integrates multiple linguistic perspectives, including Abstract Meaning Representation, constituency structure, dependency syntax, and semantic attention, augmented with external knowledge. CMV-Fuse employs a hierarchical gated fusion architecture to align local syntactic, intermediate semantic, and global knowledge representations, while a structure-aware multi-view contrastive learning objective enforces cross-view consistency without introducing additional model complexity. Experiments on three benchmark datasets demonstrate that CMV-Fuse achieves consistent and competitive performance over strong recent baselines, with analysis showing how complementary linguistic views contribute to more robust aspect-opinion reasoning.

Paper Type: Long

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

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

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 6483