MetaPrompting: Learning to Learn Better Prompts

MetaPrompting: Learning to Learn Better Prompts Yutai Hou*, Hongyuan Dong*, Xinghao Wang, Bohan Li, Wanxiang Che† Research Center for Social Computing and Information Retrieval Harbin Institute of Technology, China {ythou, hydong, xhwang, bhli, car}@ir.hit.edu.cn Abstract Prompting method is regarded as one of the cru- cial progress for few-shot nature language pro- cessing. Recent research on prompting moves from discrete tokens based “hard prompts” to continuous “soft prompts”, which employ learnable vectors as pseudo prompt tokens and achieve better performance. Though show- ing promising prospects, these soft-prompting methods are observed to rely heavily on good initialization to take effect. Unfortunately, ob- taining a perfect initialization for soft prompts requires understanding of inner language mod- els working and elaborate design, which is no easy task and has to restart from scratch for each new task. To remedy this, we pro- pose a generalized soft prompting method called MetaPrompting, which adopts the well- recognized model-agnostic meta-learning algo- rithm to automatically find better prompt ini- tialization that facilitates fast adaptation to new prompting tasks. Extensive experiments show MetaPrompting tackles soft prompt initializa- tion problem and brings significant improve- ment on four different datasets (over 7 points improvement in accuracy for 1-shot setting), achieving new state-of-the-art performance. 1 Introduction Enabling models to learn from a few labeled ex- amples, i.e., Few-Shot Learning (FSL), is one of the key steps toward more human-like artificial intelligence. Recently, taking advantage of large- scale Pretrained Language Models (PLM) (Brown et al., 2020), prompting-based methods achieve im- pressive results for few-shot learning of Natural Language Processing (NLP) (Gao et al., 2021; Liu et al., 2021a; Zhao et al., 2021). Prompting-based methods insert a piece of text, i.e. prompts, to the input examples, so that the few- shot task can be formulated as a (masked) language * Equal contribution. †Email corresponding Figure 1: Comparison between conventional soft- prompting method (left) and proposed MetaPrompting (right). x denotes the query sentence, and z is learnable pseudo tokens in soft prompts. ϕ represents all train- able parameters. MetaPrompting exploits optimization- based meta-learning to find an initialization ϕmeta that facilitates better and faster adaptation to new tasks. modeling problem. For example, say we want to classify the sentiment of the book review “I will never read it again.”, we can append a prompt “It was” to the sentence, getting “I will never read it again. It was”. It is natural to expect a higher prob- ability from the PLM to generate “terrible” than “great” then. Such converting bridges the gap be- tween pre-training and target tasks. Consequently, it has better transferability and less dependence on target task data. The performance of prompting methods is found to be greatly affected by the design of prompts (Gao et al., 2021). That is, a good prompt makes sig- nificant difference. Early attempts take manually- designed prompts or search prompts automatically. Schick et al. (2020) and (Schick and Schütze, 2021) explore to automatically identify label words. In pursuit of better performance compared to hand- picked prompts, Gao et al. …

Content truncated to 500 words