Scalable Prompt Generation for Semi-supervised Learning with Language Models

Findings of the Association for Computational Linguistics: EACL 2023, pages 770–781 May 2-6, 2023 ©2023 Association for Computational Linguistics Scalable Prompt Generation for Semi-supervised Learning with Language Models Yuhang Zhou* University of Maryland College Park, MD tonyzhou@umd.edu Suraj Maharjan* Amazon Seattle, WA mhjsuraj@amazon.com Beiye Liu Amazon New York, NY beiyeliu@amazon.com Abstract Prompt-based learning methods in semi- supervised learning (SSL) settings have been shown to be effective on multiple natural lan- guage understanding (NLU) datasets and tasks in the literature. However, manually designing multiple prompts and verbalizers requires do- main knowledge and human effort, making it difficult and expensive to scale across different datasets. In this paper, we propose two meth- ods to automatically design multiple prompts and integrate automatic verbalizer in SSL set- tings without sacrificing performance. The first method uses various demonstration examples with learnable continuous prompt tokens to cre- ate diverse prompt models. The second method uses a varying number of soft prompt tokens to encourage language models to learn differ- ent prompts. For the verbalizer, we use the prototypical verbalizer to replace the manual one. In summary, we obtained the best average accuracy of 73.2% (a relative improvement of 2.52% over even the previous state-of-the-art SSL method with manual prompts and verbal- izers) in different few-shot learning settings. 1 Introduction Pre-training large language models with huge amounts of text corpora in masked language mod- eling tasks and then fine-tuning the pre-trained lan- guage model (PLM) on downstream tasks have shown superior performance in many natural lan- guage processing tasks. However, the discrep- ancy between the pretraining task (masked lan- guage modeling objective) and the downstream fine-tuning task (task without MASK token) could lead to unexpected behaviors. Recently, there has been growing research interest in the area of prompt-tuning, where any NLU task is transformed into a cloze task to mimic the pre-training objective of a large masked language model (Kumar et al., *Equal contribution. This work was done during Yuhang’s internship at Amazon, Alexa AI. 2016; McCann et al., 2018; Radford et al., 2018). Prompt-based learning transforms an input x into x′ using a prompt function. It makes use of the vast amount of acquired knowledge of PLMs to predict a distribution of tokens at the masked posi- tion. The verbalizer then maps the predicted tokens to classes. The main advantage of this approach is that this method works well in a few-shot learning environment (Schick and Sch¨utze, 2021). How- ever, the main disadvantage of this method is the limitation posed by the prompt and verbalizer func- tions, which require human knowledge to carefully craft them. Such handcrafting work is expensive and not scalable with the increase in the variety of tasks and datasets. For example, in Alexa, there are thousands of domains and manually designing prompts and verbalizer for intent classification for each of them according to the dataset content de- mand human expertise, which is time consuming and not applicable. It is essential to reduce the human efforts in the process of prompt generation. Prompt-based …

Content truncated to 500 words