Title: DiffuSent: Advancing Towards a Unified Diffusion Framework for Aspect-Based Sentiment Analysis

Abstract:

Aspect-Based Sentiment Analysis (ABSA) comprises seven unique subtasks, each targeting specific extracted components. Although generative models have shown significant promise in unifying aspect sentiment analysis, current methods typically depend on auto-regressive, token-by-token generation. This approach often fails to capture the holistic context of aspect and opinion terms, leading to boundary insensitivity—especially when dealing with multi-word aspects or opinions.

To overcome these limitations, we introduce DiffuSent, a non-auto-regressive diffusion framework that systematically models all ABSA subtasks as boundary denoising diffusion processes. This method progressively refines boundaries starting from noisy states. Additionally, we propose a contrastive denoising training strategy designed to mitigate duplicate predictions caused by subtle variations inherent in the diffusion process.

Our extensive evaluations across 28 configurations (covering 7 subtasks and 4 datasets) reveal that DiffuSent consistently outperforms the most advanced generative and span-based systems. The framework demonstrates significant improvements on multi-word triplets, with an average F1 score increase of +2.48. It also maintains high extraction accuracy in sentences featuring multiple sentiment triplets. Furthermore, the non-auto-regressive decoding mechanism offers substantial efficiency gains, achieving inference speeds up to 181 times faster than auto-regressive generative baselines.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC