ECTTLNER: An Effective Cross-Task Transferring Learning Method for Low-Resource Named Entity Recognition

Abstract Named entity recognition (NER) is an important task in natural language processing that affects the performance of its downstream tasks. Cross-task transfer learning (CTTLNER) is more naturally suited for low-resource NER compared to cross-language transfer learning (CLTLNER) and cross-domain transfer learning (CDTLNER). Existing CTTLNER methods improve the performance of low-resource NER by leveraging relevant information from other auxiliary tasks, such as sentence-level or token-level information. However, these methods do not fully exploit token-level information of entities, resulting in room for improvement in low-resource NER. In order to improve the performance of low-resource NER, this paper proposes a simple and effective cross-task transfer learning method ECTTLNER, which introduces Sentence Contains Entities (SCE), Sentence Entity Number (SEN), Token Is Entity (TIE), and Token Boundary Label (TBL) prediction tasks to NER and performs multi-task learning together with the main sequence labeling (SEQLAB) task. Experimental results on three benchmark datasets demonstrate that ECTTLNER outperforms a set of state-of-the-art baseline models and achieves more than 2.6% improvement of F1-score in low-resource scenarios.