Knowledgeable or Educated Guess? Revisiting Language Models as Knowledge Bases

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1860–1874 August 1–6, 2021. ©2021 Association for Computational Linguistics 1860 Knowledgeable or Educated Guess? Revisiting Language Models as Knowledge Bases Boxi Cao1,3, Hongyu Lin1∗, Xianpei Han1,2∗, Le Sun1,2 Lingyong Yan1,3, Meng Liao4, Tong Xue4, Jin Xu4 1Chinese Information Processing Laboratory 2State Key Laboratory of Computer Science Institute of Software, Chinese Academy of Sciences, Beijing, China 3University of Chinese Academy of Sciences, Beijing, China 4Data Quality Team, WeChat, Tencent Inc., China {boxi2020,hongyu,xianpei,sunle,lingyong2014}@iscas.ac.cn {maricoliao,xavierxue,jinxxu}@tencent.com Abstract Previous literatures show that pre-trained masked language models (MLMs) such as BERT can achieve competitive factual knowl- edge extraction performance on some datasets, indicating that MLMs can potentially be a reli- able knowledge source. In this paper, we con- duct a rigorous study to explore the underly- ing predicting mechanisms of MLMs over dif- ferent extraction paradigms. By investigating the behaviors of MLMs, we find that previous decent performance mainly owes to the biased prompts which overfit dataset artifacts. Fur- thermore, incorporating illustrative cases and external contexts improve knowledge predic- tion mainly due to entity type guidance and golden answer leakage. Our findings shed light on the underlying predicting mechanisms of MLMs, and strongly question the previous conclusion that current MLMs can potentially serve as reliable factual knowledge bases1. 1 Introduction Recently, pre-trained language models (Peters et al., 2018; Devlin et al., 2019; Brown et al., 2020) have achieved promising performance on many NLP tasks. Apart from utilizing the universal representa- tions from pre-trained models in downstream tasks, some literatures have shown the potential of pre- trained masked language models (e.g., BERT (De- vlin et al., 2019) and RoBERTa (Liu et al., 2019b)) to be factual knowledge bases (Petroni et al., 2019; Bouraoui et al., 2020; Jiang et al., 2020b; Shin et al., 2020; Jiang et al., 2020a; Wang et al., 2020; Kass- ner and Sch¨utze, 2020a; Kassner et al., 2020). For example, to extract the birthplace of Steve Jobs, we can query MLMs like BERT with “Steve Jobs was born in [MASK]”, where Steve Jobs is the subject ∗Corresponding Authors 1We openly release the source code and data at https: //github.com/c-box/LANKA Prompt Bias “was born in” without X predicts <?> Type Guidance <?> will have the same type as B Answer Leakage Context helps if it leaks <?> Mechanism Prompt-based X was born in <?>. Case-based A was born in B. X was born in <?>. Context-based X lives in Y. X was born in <?>. Paradigm Figure 1: This paper explores three different kinds of factual knowledge extraction paradigms from MLMs, and reveal the underlying predicting mechanisms be- hind them. of the fact, “was born in” is a prompt string for the relation “place-of-birth” and [MASK] is a placeholder for the object to predict. Then MLMs are expected to predict the correct answer “Califor- nia” at the [MASK] position based on its internal knowledge. To help MLMs better extract knowl- edge, the …

Content truncated to 500 words