A prototype clustering network for enhancing semantic differentiation in named entity recognition

In named entity recognition, all spans of a sentence can be organized into a two-dimensional (2D) representation for resolving semantic structures of a sentence. Representing overlapped entities and learning the semantic dependency between them is effective. An important phenomenon in the 2D representation is that adjacent spans correspond to overlapped phrases in a sentence. Therefore, the semantics of a true entity span is shared by adjacent spans, which leads to a serious semantic diffusion problem. In related works, adjacent spans are roughly suppressed in the optimization process, leading to an unreconciled situation when tokens are simultaneously shared by entity and non-entity spans. In this study, a prototype clustering network (PCN) is proposed to utilize the semantic expression in the 2D representation. The network maps all entity types into a set of prototypes. Then, a clustering strategy is adopted to leverage type information for capturing the intra-class diversity, ensuring that adjacent spans are matched to the closest prototype clusters. It enhances the differentiation between adjacent spans in similar contexts. Experimental results demonstrate that our method achieves competitive performance on three public datasets. Analytical experiments are conducted to evaluate the clustering mechanism on the 2D representation for reducing the semantic diffusion problem, underscoring the effectiveness of our method in named entity recognition. Furthermore, it shows that the potential of the PCN can be extended for supporting other natural language processing tasks.