ETRI Knowledge Sharing Platform : Sentiment Hypergraphs With Hierarchical Reasoning: A Novel Framework for Advanced Aspect-Based Sentiment Analysis

Titles

논문 검색
Type		SCI
Year	~	Keyword

List

Journal Article Sentiment Hypergraphs With Hierarchical Reasoning: A Novel Framework for Advanced Aspect-Based Sentiment Analysis

Cited 0 time in scopus

Authors: Anusha P. Anilkumar, Yeo-Chan Yoon, Kyu-Hyung Kim, Dae-Heon Park, Ki-Duck Kim

Citation: IEEE Transactions on Affective Computing, v.17, no.3, pp.3461-3478

Abstract: We introduce SHHR (Sentiment Hypergraphs with Hierarchical Reasoning), a novel framework that unifies structural modeling, semantic enrichment, and reasoning-based inference for Aspect-Based Sentiment Analysis. Unlike conventional approaches that rely on shallow pairwise attention or static syntactic trees, SHHR constructs dynamic hypergraphs to encode higher-order dependencies between aspect terms, opinion tokens, and latent sentiment paths. These graphs are further infused with external knowledge from ConceptNet and WordNet, forming hybrid structures that anchor sentiment interpretation in both contextual and commonsense understanding. A multi-hop reasoning path mechanism traverses these fused graphs to aggregate semantically interpretable reasoning paths with component-level attribution, while a hierarchical attention module seamlessly combines local and global context. Together, these components enable SHHR to perform sentiment propagation over both structurally ambiguous and semantically rich utterances. Empirical results on SemEval-2014 (Laptop and Restaurant) and the MAMS benchmark demonstrate consistent improvements over leading transformer-based, graph-based, and knowledge-augmented ABSA models. Ablation studies, statistical tests, and detailed error analysis highlight the robustness and interpretability of SHHR, affirming its utility in modeling complex sentiment flows beyond surface-level cues.

Keyword: Aspect-based sentiment analysis, attention mechanisms, graph neural networks, hierarchical reasoning, knowledge graph fusion, multi-hop reasoning, natural language processing, sentiment hypergraphs

KSP Keywords: Attention mechanism, External knowledge, Global context, Graph fusion, Graph-based, Hierarchical reasoning, Higher-order, Hybrid structure, Multi-Hop, Natural Language processing, Neural network