Enhancing NER with Sentence-Level Entity Detection as an Simple Auxiliary Task

Chen Wang, Cong Hu, Jiang Zhong, Huawen Liu, Qi Li, Donghua Yu, Xue Li

Abstract: Named Entity Recognition (NER) is a crucial task in natural language processing (NLP) that identifies specific entities within unstructured text. However, NER models are traditionally reliant on extensive manual annotations, which is both laborious and costly. To address this challenge, we propose a simple yet effective multi-task learning framework that requires no additional labeling efforts. Our approach leverages the observation that nearly 35%–45% sentences of the existing datasets do not contain any entities. In specific, we introduce a sentence-level entity detection auxiliary task to enrich the primary NER task. The label for the auxiliary task could be directly inferred from the NER labels. This dual-task strategy not only enhances model performance but also represents good generalization over multiple NER datasets. Our experiments on the MSRA and Weibo NER datasets show that our method could effectively boost the existing state-of-the-art NER methods, offering a compelling avenue for the advancement of efficient and robust NER methods.