Does Pre-training Induce Systematic Inference? How Masked Language Models Acquire Commonsense Knowledge

TL;DR

This study investigates whether masked language models like BERT acquire commonsense knowledge through systematic inference or surface-level patterns, finding that they primarily learn from co-occurrence rather than reasoning.

Contribution

The paper introduces a method to test if pre-training induces systematic inference, revealing that models mainly learn from surface patterns, not reasoning.

Findings

Generalization does not improve during pre-training.

Commonsense knowledge is mainly from surface-level patterns.

Systematic inference is not significantly induced by pre-training.

Abstract

Transformer models pre-trained with a masked-language-modeling objective (e.g., BERT) encode commonsense knowledge as evidenced by behavioral probes; however, the extent to which this knowledge is acquired by systematic inference over the semantics of the pre-training corpora is an open question. To answer this question, we selectively inject verbalized knowledge into the minibatches of a BERT model during pre-training and evaluate how well the model generalizes to supported inferences. We find generalization does not improve over the course of pre-training, suggesting that commonsense knowledge is acquired from surface-level, co-occurrence patterns rather than induced, systematic reasoning.