Discussion of Moves Recognition of Scientific Documents Under Limited Samples
[Purpose/Significance]Moves recognition refers to extracting semantic segments such as research purposes,objects,methods,results,and conclusions from unstructured abstracts.This paper focuses on the problems of the limited high-quality annotation samples and the poor interpretability of the deep recognition model that often occur in the practical application of move recognition.[Method/Process]In this paper,it introduced the prompt-based learn-ing paradigm in move recognition,and designed the corresponding template and verbalizer.By the local linear proxy approach,it produced the model interpretation and constructed an interpretable deep learning recognition model.Then it carried out a simulation empirical study on randomly selected partial data from two datasets in the biological and computer fields.[Result/Conclusion]Prompt tuning on large model can achieve higher accuracy than the fine-tuned small model on moves recognition task with less training cost.After training on three sub datasets of PubMed,the F1 score was improved by 2.5％,4.1％and 3.9％,respectively.Combining the accuracy rate and interpretation results,the"method"and"result"moves recognition effect is better(F1 score about 90％),and the"background"and"method"moves are relatively poor(F1 score＜70％).The prompt learning approaches are faster and more efficient to use large pre-trained language model,and obtain recognition results with high accuracy and interpretability.
万方数据
维普
