Prompt-based Few-shot Learning for Academic Document Classification
By studying the long-tail phenomenon and emerging classification problems in academic document classification,this paper proposes a few-shot document classification method based on prompt learning to achieve automatic classification in low-resource scenarios.With the capabilities of text representation and generation from large-scale pre-trained language models(PLMs),the effects of different prompt templates,document fields,classification representations,number of samples,and other factors on document classification within the prompt learning framework are analyzed.Experimental results show that,through rationally designing prompt templates,document classification representations,document fields,and others,the proposed model is able to effectively achieve document classification in low-resource scenarios with an F1 value of 85%for 50 shots,which is an important complement to traditional document classification algorithms.However,there are some limitations in fine-grained classification that need to be improved.
few-shot learningprompt learningacademic document classificationpre-trained language model
NETL
NSTL
万方数据
