Acronym identification based on SpanBERT-CRF

Acronym Identification means to recognizing acronyms and their definitions in a given paragraph by an algorithm. Existing studies have shown that the pre-training model has significant performance advantages in acronyms recognition. However, the vanilla Bert measures randomly masking tokens in the pre-training, which may also abbreviate the token of acronyms in sentence. The continuous masking method will be more helpful to the recognition and interpretation of acronyms. In this paper, we propose to use the SpanBERT model for acronym recognition and obtaining the transfer relationship between indicators with the help of a conditional random field (CRF). Through experimental comparison, the results demonstrate that our method increases the F1 value compared with the Bert and BERT-CRF models.