Are Pretrained Language Models Symbolic Reasoners Over Knowledge?
This work addresses the fundamental problem of knowledge acquisition in PLMs for AI researchers, providing insights into their limitations in symbolic reasoning.
The study investigates how pretrained language models (PLMs) learn factual knowledge, focusing on reasoning and memorization mechanisms, and finds that PLMs partially learn symbolic reasoning rules but struggle with complex tasks like two-hop reasoning, while memorization success depends on schema conformity and frequency.
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.