DVDNER: Dual-View Learning Named Entity Recognition via Diffusion

Named Entity Recognition (NER) is a critical component in the construction of knowledge graphs. However, traditional approaches often neglect the discrepancy between training and evaluation environments, leading to suboptimal model performance. To tackle this, we propose DVDNER (Dual-view Learning Named Entity Recognition via Diffusion), a novel approach that reconceptualizes NER through a boundary-denoising diffusion paradigm. DVDNER is equipped with its unique iterative denoising training mechanism, which systematically introduces noise around entity boundaries and employs a reverse diffusion process for their accurate recovery. Specifically, DVDNER is twofold: (1) It enables the model to incorporate structured knowledge of data points during the denoising phase, maintaining a coherent transition from raw data to pure noise in the diffusion model. This strategy adeptly addresses the intricate challenges of recognizing entities in diverse contexts. (2) Rooted in a dual-view learning paradigm, DVDNER effectively merges the understanding of individual entities with their broader sentence-level contexts. This integration is key to capturing the essential contextual information for precise entity boundary recognition and effective denoising. Extensive experiments demonstrate that DVDNER achieves state-of-the-art performance compared with previous models.