GPS: Genetic Prompt Search for Efficient Few-Shot Learning

Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, pages 8162 - 8171 December 7-11, 2022 ©2022 Association for Computational Linguistics GPS: Genetic Prompt Search for Efficient Few-shot Learning Hanwei Xu∗, Yujun Chen∗, Yulun Du∗, Nan Shao, Yanggang Wang, Haiyu Li, Zhilin Yang† Recurrent AI {xuhanwei, chenyujun, duyulun, kimi_yang}@rcrai.com Abstract Prompt-based techniques have demostrated great potential for improving the few-shot generalization of pretrained language models. However, their performance heavily relies on the manual design of prompts and thus requires a lot of human efforts. In this paper, we intro- duce Genetic Prompt Search (GPS) to improve few-shot learning with prompts, which utilizes a genetic algorithm to automatically search for high-performing prompts. GPS is gradient-free and requires no update of model parameters but only a small validation set. Experiments on di- verse datasets proved the effectiveness of GPS, which outperforms manual prompts by a large margin of 2.6 points. Our method is also better than other parameter-efficient tuning methods such as prompt tuning. 1 Introduction Pretrained language models, such as BERT (Devlin et al., 2019), XLNet (Yang et al., 2019), T5 (Raffel et al., 2020), and GPT (Radford et al., 2018), are often finetuned for downstream natural language processing tasks, which has been shown to improve performance over non-pretrained models. However, this pretraining-finetuning paradigm still relies on a relatively large set of labeled data for each down- stream task to obtain competitive performance. Al- though GPT-3 (Brown et al., 2020) shows promis- ing performance for zero-shot and few-shot learn- ing by prompting on an extremely-large pretrained language models with 175B parameters, finding out the optimum prompt for each given task could be difficult (Mishra et al., 2021; Wang et al., 2022). To improve the performance of prompting on pretrained language models, recent works focus on supervised pretraining with carefully designed ∗ Equal contribution † Corresponding author Code is available at https://github.com/hwxu20/GPS or crowdsourced manual prompts (Gao et al., 2021a; Wei et al., 2021; Sanh et al., 2021; Ouyang et al., 2022). Diverse prompts are collected to en- hance the robustness and performance of prompt- ing (Sanh et al., 2021). Ouyang et al. (2022) in- troduced a dataset of labeler demonstrations and used it to finetune GPT-3. Despite all these efforts, the challenge to obtain high-performing prompts for few-shot learning still exists. As pointed out by previous works (Liu et al., 2021b; Gao et al., 2021b; Liu et al., 2021a), manual prompts are usu- ally suboptimal and suffers a high variance on per- formance. To address this challenge, we propose a novel Genetic Prompt Search (GPS) algorithm that grad- ually mutates the prompts with a generative model and selects candidates according to their perfor- mance on a small development set. This evolution- ary procedure relies on a tiny set of labeled data, only used for validation but not training. As illus- trated in Figure 1, GPS does not require updating any parameter, but only searches for the optimal hard prompts for every downstream task. Similar to prompt tuning, GPS …

Content truncated to 500 words