Improving Named Entity Recognition with Multi-Feature Word-to-Word Relationship Classification

In the realm of Named Entity Recognition (NER), three predominant forms are recognized: flat, overlapped, and discontinuous NER. The conventional approaches include Sequence Labeling, Span-based, and Sequence-to-Sequence methods, each beset with distinct challenges in terms of efficiency or efficacy. In our novel contribution, we introduce an innovative framework that augments NER performance by delineating word-to-word relationships. This is accomplished by mining multiple features and forging two distinct relationship types: Next-Adjacent-Word (NAW) and Tail-Head-Word (THW). Our framework employs multi-granularity two-dimensional convolutions alongside Multi-head Biaffine attention to capture more nuanced representations. Empirical evidence from our experiments indicates that our model achieves competitive results on seven extensively recognized benchmark datasets.