Learning to Initialize: Can Meta Learning Improve Cross-task Generalization in Prompt Tuning?

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 11802–11832 July 9-14, 2023 ©2023 Association for Computational Linguistics Learning to Initialize: Can Meta Learning Improve Cross-task Generalization in Prompt Tuning? Chengwei Qin♣, Shafiq Joty♣♠, Qian Li�, Ruochen Zhao♣ ♣ Nanyang Technological University ♠ Salesforce AI � Northeastern University {chengwei003@e.ntu, srjoty@ntu, ruochen002@e.ntu}.edu.sg qianli@stumail.neu.edu.cn Abstract Prompt tuning (PT) which only tunes the em- beddings of an additional sequence of tokens per task, keeping the pre-trained language model (PLM) frozen, has shown remarkable performance in few-shot learning. Despite this, PT has been shown to rely heavily on good initialization of the prompt embeddings. In this work, we study meta prompt tuning (MPT) to systematically explore how meta-learning can help improve (if it can) cross-task general- ization in PT through learning to initialize the prompt embeddings from other relevant tasks. We empirically analyze a representative set of meta learning algorithms in a wide range of adaptation settings with different source/target task configurations on a large set of few-shot tasks. With extensive experiments and analysis, we demonstrate the effectiveness of MPT. We find the improvement to be significant particu- larly on classification tasks. For other kinds of tasks such as question answering, we observe that while MPT can outperform PT in most cases, it does not always outperform multi-task learning. We further provide an in-depth analy- sis from the perspective of task similarity. 1 Introduction Humans can easily learn to perform new tasks with only few data by leveraging previously acquired knowledge from other relevant tasks. Such capa- bility is a hallmark of human intelligence (Carey and Bartlett, 1978). However, when it comes to the models, they often face over-fitting issues when they are tasked to learn from a few labeled exam- ples (Lake et al., 2017; Linzen, 2020), a problem commonly termed as few-shot learning (FSL). With the recent advancements in developing large-scale pre-trained language models (PLMs), prompt-based methods have shown promising re- sults in FSL. Brown et al. (2020) show that by virtue of in-context (meta) learning, a frozen GPT- 3 model can achieve good results on a variety of Upstream learning Sentiment Analysis Question Answering ... Input Summarization Downstream learning Topic Classification Input Output The best book ever to anything. Output When was Kamala Harris born? He gains the access to a fortune of $320 million... Positive October 20, 1964 ... He gets $320M fortune... Usain Bolt is the greatest sprinter of all time. Sports Source Tasks Target Task Learner Figure 1: Illustration of cross-task generalization, where the model is expected to learn an unseen target task given the knowledge acquired from previously learned source tasks. few-shot tasks through manually designed prompts, which are task instructions along with a few exam- ples expressed in natural language. However, the performance of in-context learning has been shown to be highly sensitive to the design of such “dis- crete” prompts (Zhao et al., 2021). It is also limited by the maximum sequence length supported by the PLMs …

Content truncated to 500 words