Word Pair Relation Classification Model for Nested Named Entity Recognition

Abstract

Nested named entity recognition is aimed at overlapping entities. The span-based approach considers nested named entity recognition as a two-stage task: span extraction and span classification. However, this method has some problems such as error propagation and long entity recognition. Therefore, this paper put forward a word pair relationship classification model, which describes nested entities as words and multi relationship classification tasks between words. Firstly, the grid matrix representation of word pairs is constructed by traversing the text. Then, in order to enhance the connection between word and word, the distance information and position information of word pairs are fused into the grid matrix representation of word pairs. Finally, through the grid matrix to prediction category relationships of all pairs of words. The experimental results show that the F1-score of the proposed word pair classification model reaches 77.3% on the nested entity recognition dataset GENIA, and its nested entity recognition ability is superior to the existing models.