A Boundary Feature Enhanced Span-Based Nested Named Entity Recognition Method

Named Entity Recognition (NER) is a crucial task in natural language processing. Traditional NER methods mainly focus on identifying flat named entities and show poor performance in recognizing Nested Named Entities (NNEs). Towards Nested Named Entity Recognition (NNER), span-based methods, as a mainstream, have been proposed recently. The effectiveness of identifying entity span, which can be regarded as sub-sequences of a sentence, directly affects the performance of this kind of methods. However, the identification of irrelevant entity spans remains a challenge in academia. In this paper, we jointly take contextual semantics and entity boundaries into account and propose a novel Span-based NNER method (BFSN2ER). In BFSN2ER, BERT and Bi-LSTM are integrated to extract contextual semantics of entities, which is further used to detect the beginning and end boundaries of entities. Moreover, to improve the efficiency of BFSN2ER, we introduce a multi-task learning framework to achieve jointly models training. To validate the performance of BFSN2ER, experiments were conducted on three large datasets. Comparing with seven baselines, BFSN2ER achieved obviously better recall and F1-score, which indicates that BFSN2ER can recognize more accurate NNEs.