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

Researchers introduce DiffuSent, a non-autoregressive diffusion framework that reformulates seven aspect-based sentiment analysis (ABSA) subtasks as boundary denoising processes. The approach achieves significant improvements over existing generative models, particularly on multi-word expressions, while delivering up to 181x faster inference speeds through parallel decoding rather than sequential token generation.

DiffuSent addresses a fundamental limitation in current sentiment analysis systems: their inability to accurately identify boundaries of multi-word aspect and opinion terms. Traditional autoregressive approaches generate tokens sequentially, which often results in imprecise extraction and duplicate predictions with minor variations. By reframing ABSA subtasks as diffusion processes that progressively denoise boundary predictions, the framework fundamentally changes how sentiment information is extracted from text.

The research builds on the growing momentum of diffusion models beyond image generation into NLP tasks. While generative models have proven effective for unified ABSA, they struggle with the granular precision required for boundary detection. DiffuSent's introduction of contrastive denoising training further refines the approach by reducing spurious duplicates—a persistent problem in diffusion-based NLP systems. The experimental validation across 28 settings (7 subtasks across 4 datasets) provides robust evidence of the framework's effectiveness.

From a practical standpoint, the 181x inference speedup holds significant implications for deployment at scale. Organizations running sentiment analysis systems can dramatically reduce computational costs while improving accuracy, making sentiment-driven applications more economically viable. The particularly strong performance on multi-word triplets suggests DiffuSent excels at handling complex, real-world text where sentiment expressions span multiple tokens—a common occurrence in product reviews, financial sentiment, and social media analysis.

Future developments likely involve extending diffusion frameworks to other structured NLP tasks beyond sentiment analysis, potentially creating a unified paradigm for information extraction problems that require precise boundary detection.

• →DiffuSent achieves +2.48 average F1 improvement on multi-word sentiment triplets compared to existing generative and span-based systems.
• →Non-autoregressive decoding enables 181x faster inference than autoregressive baselines through parallel boundary prediction.
• →Contrastive denoising training effectively eliminates duplicate predictions, a major limitation of previous diffusion-based NLP approaches.
• →The framework unifies seven distinct ABSA subtasks under a single diffusion-based methodology, simplifying model architecture and training.
• →Performance gains are robust across 28 experimental settings, demonstrating consistent improvements on multiple benchmark datasets.