AutoPrompt: Eliciting Knowledge from Language Models with Automatically Generated Prompts

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4222–4235, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4222 AUTOPROMPT: Eliciting Knowledge from Language Models with Automatically Generated Prompts Taylor Shin∗♦ Yasaman Razeghi∗♦ Robert L. Logan IV∗♦ Eric Wallace♠ Sameer Singh♦ ♦University of California, Irvine ♠University of California, Berkeley {tshin1, yrazeghi, rlogan, sameer}@uci.edu ericwallace@berkeley.edu Abstract The remarkable success of pretrained lan- guage models has motivated the study of what kinds of knowledge these models learn dur- ing pretraining. Reformulating tasks as fill- in-the-blanks problems (e.g., cloze tests) is a natural approach for gauging such knowledge, however, its usage is limited by the manual effort and guesswork required to write suit- able prompts. To address this, we develop AUTOPROMPT, an automated method to cre- ate prompts for a diverse set of tasks, based on a gradient-guided search. Using AUTO- PROMPT, we show that masked language mod- els (MLMs) have an inherent capability to perform sentiment analysis and natural lan- guage inference without additional parame- ters or finetuning, sometimes achieving per- formance on par with recent state-of-the-art supervised models. We also show that our prompts elicit more accurate factual knowl- edge from MLMs than the manually created prompts on the LAMA benchmark, and that MLMs can be used as relation extractors more effectively than supervised relation extraction models. These results demonstrate that au- tomatically generated prompts are a viable parameter-free alternative to existing probing methods, and as pretrained LMs become more sophisticated and capable, potentially a re- placement for finetuning. 1 Introduction Pretrained language models (LMs) have had ex- ceptional success when adapted to downstream tasks via finetuning (Peters et al., 2018; Devlin et al., 2019). Although it is clear that pretrain- ing improves accuracy, it is difficult to determine whether the knowledge that finetuned LMs contain is learned during the pretraining or the finetuning process. How can we directly evaluate the knowl- ∗First three authors contributed equally. edge present in pretrained LMs, be it linguistic, factual, commonsense, or task-specific? Numerous techniques have been proposed to elicit such knowledge by analyzing pretrained LMs’ internal representations. A common strategy is to use probing classifiers—shallow classifiers that predict certain attributes using an LMs’ representa- tions as features (Conneau et al., 2018; Liu et al., 2019). However, probing classifiers require ad- ditional learned parameters and are thus suscep- tible to false positives; high probing accuracy is not a sufficient condition to conclude that an LM contains a certain piece of knowledge (Hewitt and Liang, 2019; Voita and Titov, 2020). Attention visualization, another common technique, has a similar failure mode: attention scores may be corre- lated with, but not caused by the underlying target knowledge, leading to criticism against their use as explanations (Jain and Wallace, 2019; Wiegreffe and Pinter, 2019). Both probing and attention visu- alizations also struggle to evaluate knowledge that cannot be represented as simple token- or sequence- level classification tasks. A more direct approach for eliciting knowledge from these models, since they are language models after all, is …

Content truncated to 500 words