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

Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies, pages 5017–5033 June 6–11, 2021. ©2021 Association for Computational Linguistics 5017 Factual Probing Is [MASK]: Learning vs. Learning to Recall Zexuan Zhong∗ Dan Friedman∗ Danqi Chen Department of Computer Science Princeton University {zzhong, dfriedman, danqic}@cs.princeton.edu Abstract Petroni et al. (2019) demonstrated that it is possible to retrieve world facts from a pre- trained 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. Sub- sequent work has attempted to tighten the es- timate by searching for better prompts, using a disjoint set of facts as training data. In this work, we make two complementary contribu- tions to better understand these factual probing techniques. First, we propose OPTIPROMPT, a novel and efficient method which directly op- timizes in continuous embedding space. We find this simple method is able to predict an additional 6.4% of facts in the LAMA bench- mark. Second, we raise a more important ques- tion: Can we really interpret these probing re- sults 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 con- tains certain regularities of the underlying fact distribution, and all the existing prompt meth- ods, 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 de- tailed picture of what different prompts can re- veal about pre-trained language models.1 1 Introduction Pre-trained language models like BERT are op- timized to predict the distribution of words in an Internet corpus (Devlin et al., 2019). Naturally, this distribution encodes information about world facts. Recently, researchers have taken an interest in mea- suring how much factual information language models acquire from pre-training. Petroni et al. (2019) formally define this project in the LAMA *The first two authors contributed equally. 1The code is publicly available at https://github. com/princeton-nlp/OptiPrompt. probe: X was born in __ BERT Linguistic probe: dependency label The chef made pizzas probe: classifier nsubj Factual probe: (X, place_of_birth, ?) John Milton London Training data: Training data: (Dante, place_of_birth, Florence) (John Donne, place_of_birth, London) (Guy Deghy, place_of_birth, Budapest) … Alice saw Bob nsubj dobj BERT Figure 1: A linguistic probe is trained to predict linguis- tic annotations given the representations returned by a language model, and evaluated on a held-out set of sen- tences. A factual probe is trained to predict an object for a subject and a relation using a pre-trained language model, and evaluated on a held-out set of subject-object pairs that express the same relation. benchmark, which consists of (subject, relation, object) triples along with human-written templates that express each relation. They show that BERT can predict objects given cloze-style prompts—for example, “Dante was born in [MASK]”—and they present their result as …

Content truncated to 500 words