Knowledgeable or Educated Guess? Revisiting Language Models as Knowledge Bases

TL;DR

The paper critically reexamines whether pre-trained masked language models can function as factual knowledge bases by analyzing three extraction paradigms. It demonstrates that prompt-based success largely results from prompt-induced biases and dataset overfitting, that case-based illustrations mainly aid object-type recognition rather than exact entity identification, and that context-based improvements largely rely on explicit or implicit answer leakage. These findings collectively challenge the view that MLMs reliably encode factual knowledge, and they illuminate the mechanisms by which different paradigms influence predictions. The work also proposes methods for better understanding and evaluating knowledge extraction from MLMs and suggests directions for future research, including extending analysis to generative LMs and refining evaluation strategies.

Abstract

Previous literatures show that pre-trained masked language models (MLMs) such as BERT can achieve competitive factual knowledge extraction performance on some datasets, indicating that MLMs can potentially be a reliable knowledge source. In this paper, we conduct a rigorous study to explore the underlying predicting mechanisms of MLMs over different extraction paradigms. By investigating the behaviors of MLMs, we find that previous decent performance mainly owes to the biased prompts which overfit dataset artifacts. Furthermore, incorporating illustrative cases and external contexts improve knowledge prediction 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 bases.

Figures (8)

• Figure 1: This paper explores three different kinds of factual knowledge extraction paradigms from MLMs, and reveal the underlying predicting mechanisms behind them.
• Figure 2: An illustration example of the vastly different answer distributions but similar prediction distributions on LAMA and WIKI-UNI on "place-of-birth" relation.
• Figure 3: Correlations of the prediction distributions on LAMA and WIKI-UNI. Even these two datasets have totally different answer distributions, MLMs still make highly correlated predictions.
• Figure 4: Correlations between the prompt-only distribution and prediction distribution on WIKI-UNI. MLMs make correlated predictions w. or w/o. subjects.
• Figure 5: Illustration of our type induction algorithm. The numbers on the right of each type indicate how many entities does the type cover. The type of an entity set is the finest grained type in the type graph that can cover a sufficient number of the instances in the entity set, which is City in the example.