Demystifying Prompts in Language Models via Perplexity Estimation

By Hila Gonen, Srini Iyer, Terra Blevins, Noah A. Smith, Luke S. Zettlemoyer. Conference on Empirical Methods in Natural Language Processing, 2022

313 citations recorded by the source corpus.

Language models can be prompted to perform a wide variety of zero- and few-shot learning problems. However, performance varies significantly with the choice of prompt, and we do not yet understand why this happens or how to pick the best prompts. In this work, we analyze the factors that contribute to this variance and establish a new empirical hypothesis: the performance of a prompt is coupled with the extent to which the model is familiar with the language it contains. Over a wide range of tasks, we show that the lower the perplexity of the prompt is, the better the prompt is able to perform the task. As a result, we devise a method for creating prompts: (1) automatically extend a small seed set of manually written prompts by paraphrasing using GPT3 and backtranslation and (2) choose the lowest perplexity prompts to get significant gains in performance.

View Paper: https://www.semanticscholar.org/paper/d03a9b2a0e090cc9fd2ba0a457ecea35372f1018

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

1. Ce Zheng
2. Jingjing Xu
3. Lei Li
4. Xu Sun
5. Qingxiu Dong
6. Lei Li
7. Zhiyong Wu
8. Yejin Choi

Citations in the corpus

• A Survey for In-context Learning (2023, background): “Perplex-ity (PPL) is another commonly-used metric, which computes the sentence perplexity of the whole input sequence S j = { C, s ( x, y j , I ) } consists of the tokens of demonstration examples C , input query x and target label y j .”
• A Survey on In-context Learning (2022, background): “…et al., 2022a), PICL (Gu et al., 2023) Inference (§5) Organization (§5.1) Selecting(§5.1.1) KATE (Liu et al., 2022), EPR (Rubin et al., 2022), PPL (Gonen et al., 2022), SG-ICL (Kim et al., 2022a), Self Adaptive (Wu et al., 2022),MI (Sorensen et al., 2022), Q-Learning (Zhang et al., 2022a),…”
• Quantifying Language Models' Sensitivity to Spurious Features in Prompt Design or: How I learned to start worrying about prompt formatting (2023, background): “Some existing work has explored the influence of certain prompt design choices on model performance, for example the prompt’s language (Gonen et al., 2022), the ordering of few-shot examples (Lu et al., 2022), and their patterns (Madaan et al., 2023).”
• Security and Privacy Challenges of Large Language Models: A Survey (2024, unknown): “Consequently, a data prompt is considered compromised if its perplexity exceeds a specified threshold [147].”