DuST: Chinese NER using dual-grained syntax-aware transformer network

Abstract

Recent studies have attempted to exploit syntactic information (e.g., dependency relation) to enhance Chinese named entity recognition (NER) performance and achieved promising results. These methods usually leverage single-grained syntactic parsing results which are based on single-grained word segmentation. However, entities may be annotated with varying granularities, resulting in inconsistent boundaries when compared to single-grained results. Therefore, merely using single-grained syntactic information may inadvertently introduce noise into boundary detection in Chinese NER. In this paper, we introduce a Dual-grained Syntax-aware Transformer network (DuST) to mitigate the noise introduced by single-grained syntactic parsing results. We first introduce coarse- and fine-grained syntactic dependency parsing results to comprehensively consider possible boundary scenarios. We then design the DuST network with dual syntax-aware Transformers to capture syntax-enhanced features at different granularities, a contextual Transformer to model the contextual features and an aggregation module to dynamically aggregate these features. Experiments are conducted on four widely-used Chinese NER datasets and our model achieves superior performance. Specifically, our approach outperforms two single-grained syntax-enhanced baselines with an increase of up to 3.9% and 2.94% in F1 score, respectively.