DPWE: Unified NER Model Based on Dependency Parsing and Word-Word Label Expansion
Yaodi Liu, Kun Zhang, Rong Tong, Hui Chen, Chenxi Cai, Dianying Chen · 2025
According to the entity structure, named entities can be divided into flat, nested, and discontinuous. However, in previous studies, most used sequence labeling or span classification methods to identify different types of entities separately. Still, they could not identify these three entities simultaneously, inconveniencing practical applications. Thanks to the design flexibility and better performance of the word-word relationship prediction method, in this paper, we follow this method and propose a unified named entity recognition model based on Dependency Parsing and Word-Word label Expansion (DPWE). On the one hand, inspired by the Next-Neighboring-Word (NNW) and Tail-Head-Word-* (THW-*) in the State-of-the-Art (SOTA) model, we added two new labels to the word-word prediction method, namely Front-Neighboring-Word (FNW) and Head-Tai-Word-* (HTW-*). We modeled the word-word relationship prediction as an adjacent and head-tail relationship prediction to establish a more fine-grained word-word relationship, thereby reducing the error propagation of relationship prediction. On the other hand, considering that syntactic information can effectively enhance character representation, we also use attention guided graph convolutional networks to improve our model. Experimental results show that the DPWE model proposed in this paper can achieve better recognition performance, with F values of 93.54%, 82.68%, and 74.05% on the CoNLL2003, GENIA, and CADEC datasets.