Later Span Adaptation for Language Understanding

Pre-trained contextualized language models (PrLMs) broadly use fine-grained tokens (words or sub-words) as minimal linguistic unit in pre-training phase. Introducing span-level information in pre-training has shown capable of further enhancing PrLMs. However, such methods require enormous resources and are lack of adaptivity due to huge computational requirement from pre-training. Instead of too early fixing the linguistic unit input as nearly all previous work did, we propose a novel method that combines span-level information into the representations generated by PrLMs during fine-tuning phase for better flexibility. In this way, the modeling procedure of span-level texts can be more adaptive to different downstream tasks. In detail, we divide the sentence into several spans according to the segmentation generated by a pre-sampled dictionary. Based on the sub-token-level representation provided by PrLMs, we enhance the connection between the tokens in each span and gain a representation with enhanced span-level information. Experiments are conducted on GLUE benchmark and prove that our approach could remarkably enhance the performance of PrLMs in various natural language understanding tasks.