Are Pretrained Language Models Symbolic Reasoners Over Knowledge?

TL;DR

This paper investigates how pretrained language models acquire factual knowledge, focusing on reasoning and memorization, revealing their strengths and limitations in applying reasoning rules and memorizing facts.

Contribution

It provides the first causal analysis of the relationship between training facts and learned knowledge in PLMs using synthetic data.

Findings

PLMs can apply some symbolic reasoning rules correctly

PLMs struggle with two-hop reasoning

Memorization depends on schema conformity and frequency

Abstract

How can pretrained language models (PLMs) learn factual knowledge from the training set? We investigate the two most important mechanisms: reasoning and memorization. Prior work has attempted to quantify the number of facts PLMs learn, but we present, using synthetic data, the first study that investigates the causal relation between facts present in training and facts learned by the PLM. For reasoning, we show that PLMs seem to learn to apply some symbolic reasoning rules correctly but struggle with others, including two-hop reasoning. Further analysis suggests that even the application of learned reasoning rules is flawed. For memorization, we identify schema conformity (facts systematically supported by other facts) and frequency as key factors for its success.