Co-occurrence is not Factual Association in Language Models

TL;DR

This paper reveals that language models primarily learn co-occurrence patterns rather than true factual associations, and proposes strategies to enhance factual knowledge learning and generalization in these models.

Contribution

It identifies the layer-specific encoding of knowledge in language models and introduces methods to promote learning of factual associations over co-occurrence biases.

Findings

Training on implicit factual associations improves generalization.

Forgetting co-occurrence statistics enhances factual learning.

Strategies improve reasoning performance on synthetic and real data.

Abstract

Pretrained language models can encode a large amount of knowledge and utilize it for various reasoning tasks, yet they can still struggle to learn novel factual knowledge effectively from finetuning on limited textual demonstrations. In this work, we show that the reason for this deficiency is that language models are biased to learn word co-occurrence statistics instead of true factual associations. We identify the differences between two forms of knowledge representation in language models: knowledge in the form of co-occurrence statistics is encoded in the middle layers of the transformer model and does not generalize well to reasoning scenarios beyond simple question answering, while true factual associations are encoded in the lower layers and can be freely utilized in various reasoning tasks. Based on these observations, we propose two strategies to improve the learning of factual associations in language models. We show that training on text with implicit rather than explicit factual associations can force the model to learn factual associations instead of co-occurrence statistics, significantly improving the generalization of newly learned knowledge. We also propose a simple training method to actively forget the learned co-occurrence statistics, which unblocks and enhances the learning of factual associations when training on plain narrative text. On both synthetic and real-world corpora, the two proposed strategies improve the generalization of the knowledge learned during finetuning to reasoning scenarios such as indirect and multi-hop question answering.