HIC: Named Entity Recognition Based on Enhanced Boundary Detection and Span Classification

Abstract:
The traditional named entity recognition based on sequence labeling can only assign one label to each character during prediction, which cannot solve the case of nested entities. This chapter starts from the idea of solving nested entities, and introduces a dual affine attention mechanism in the boundary detection module, so that the head and tail pairs can interact and enhance the boundary of the entity explicitly. At the same time, combined with the traditional sequence labeling model, the internal character relationship between the head and tail pairs is modeled to avoid the influence of non-entity words between the head and tail pairs. Finally, the detected head-to-tail pairs are classified by the span classification module.