Search papers, labs, and topics across Lattice.
The paper introduces GHI, a Graphormer-based framework for aspect-based sentiment analysis (ABSA) that uses a conditioned hypergraph incidence structure to represent token-hyperedge relationships. This allows the model to incorporate diverse linguistic and semantic cues through a unified interface, improving fine-grained structural reasoning. Experiments on six ABSA benchmarks show GHI outperforms baselines, demonstrating stable improvements over DeBERTa and approaching the performance of much larger Flan-T5 models with significantly fewer parameters, while also exhibiting strong robustness on challenging datasets.
Forget scaling laws for fine-grained sentiment analysis: a compact Graphormer-based model rivals 11B parameter behemoths by explicitly reasoning about token-hyperedge relationships.
Aspect-based sentiment analysis (ABSA) requires models to bind sentiment evidence to the correct aspect, making it a natural testbed for fine-grained structural reasoning. We introduce GHI, a Graphormer-over-Conditioned-Hypergraph-Incidence framework that is designed as an incidence-based structural reasoning layer built on a bipartite topology. GHI represents diverse linguistic and semantic evidence as token--hyperedge incidence relations, allowing different structural signals to be incorporated through a unified interface. Extensive experiments on six standard ABSA benchmarks show that GHI outperforms all baselines on the SemEval domains, and multi-seed evaluations show stable improvements over strong DeBERTa. Further experiments show that with only 247M parameters, GHI approaches the performance of 11B Flan-T5 based methods on the ISE benchmark. Moreover, it demonstrates strong robustness on the challenging ARTS datasets, maintaining highly competitive performance where traditional models degrade. These results demonstrate that compact structural reasoning remains a valuable alternative to scale-driven approaches for fine-grained tasks.
