Factual Probing Is [MASK]: Learning vs. Learning to Recall

Abstract

Petroni et al. (2019) demonstrated that it is possible to retrieve world facts from a pretrained language model by expressing them as cloze-style prompts and interpret the model's prediction accuracy as a lower bound on the amount of factual information it encodes. Subsequent work has attempted to tighten the estimate by searching for better prompts, using a disjoint set of facts as training data. In this work, we make two complementary contributions to better understand these factual probing techniques. First, we propose OPTIPROMPT, a novel and efficient method which directly optimizes in continuous embedding space. We find this simple method is able to predict an additional 6.4% of facts in the LAMA benchmark. Second, we raise a more important question: Can we really interpret these probing results as a lower bound? Is it possible that these prompt-search methods learn from the training data too? We find, somewhat surprisingly, that the training data used by these methods contains certain regularities of the underlying fact distribution, and all the existing prompt methods, including ours, are able to exploit them for better fact prediction. We conduct a set of control experiments to disentangle "learning" from "learning to recall", providing a more detailed picture of what different prompts can reveal about pre-trained language models. 1 * The first two authors contributed equally.

Relational World Knowledge Representation in Contextual Language Models: A Review

“…The answer is simple. Typical language Prompt handcrafting (Petroni et al, 2019; Automatic prompt engineering (Jiang et al, 2020b;Shin et al, 2020;Zhong et al, 2021;Qin and Eisner, 2021) Adversarial prompt modification Poerner et al, 2020; Varying base prompts (Elazar et al, 2021;Heinzerling and Inui, 2021;Jiang et al, 2020a; Symbolic rule-based prompting Talmor et al, 2020a) Statement scores ( § 3.2) Single-LM scoring (Tamborrino et al, 2020;) Dual-LM scoring (Davison et al, 2019;Shwartz et al, 2020) modeling corpora like Wikipedia are known to contain KB-like assertions about the world (Da and Kasai, 2019). LMs trained on enough such data can be expected to acquire some KB-like knowledge, even without targeted entity-or relation-level supervision.…”

“…Automatic prompt engineering is a promising alternative to prompt handcrafting for knowledge extraction in LMs (Liu et al, 2021a), as prompts engineered using discrete (Jiang et al, 2020b;Shin et al, 2020;Haviv et al, 2021) and continuous (Zhong et al, 2021;Qin and Eisner, 2021;Liu et al, 2021b) optimization have improved LMs' lower-bound performance on LAMA's underlying queries. Note, however, that optimized prompts are not always grammatical or intelligible (Shin et al, 2020).…”

“…Note, however, that optimized prompts are not always grammatical or intelligible (Shin et al, 2020). Prompt optimization methods may also confound knowledge probes by overfitting to the probes' answer distributions during train-ing (Zhong et al, 2021;, and often require large validation sets for tuning, which may not be feasible in practice (Perez et al, 2021).…”

On Transferability of Prompt Tuning for Natural Language Processing

“…Recently, many prompt-based works have emerged, i.e., manuallydesigned (Schick and Schütze, 2021a,b;Mishra et al, 2021) or automatically-searched (Jiang et al, 2020;Shin et al, 2020;Gao et al, 2021) hard prompts, which are discrete tokens but not necessarily human-readable. Furthermore, soft prompt (Li and Liang, 2021;Hambardzumyan et al, 2021;Zhong et al, 2021;Liu et al, 2021) comes out, which are tuneable embeddings rather than tokens in the vocabularies and can be directly trained with task-specific supervision. And demonstrates that this prompt tuning (PT) method can match the performance of full-parameter finetuning when the PLM size is extremely large.…”

Few-shot Instruction Prompts for Pretrained Language Models to Detect Social Biases

“…We harness the knowledge present in large scale pre-trained language models (Davison et al, 2019;Zhou et al, 2020;Petroni et al, 2019;Zhong et al, 2021;Shin et al, 2020) to detect a rich set of biases. Our method prompts the LM with a textual post and labeled exemplars selected using a novel technique along with instructions to detect bias in this post.…”