Can Pretrained Language Models (Yet) Reason Deductively?

Abstract

Acquiring factual knowledge with Pretrained Language Models (PLMs) has attracted increasing attention, showing promising performance in many knowledge-intensive tasks. Their good performance has led the community to believe that the models do possess a modicum of reasoning competence rather than merely memorising the knowledge. In this paper, we conduct a comprehensive evaluation of the learnable deductive (also known as explicit) reasoning capability of PLMs. Through a series of controlled experiments, we posit two main findings. (i) PLMs inadequately generalise learned logic rules and perform inconsistently against simple adversarial surface form edits. (ii) While the deductive reasoning fine-tuning of PLMs does improve their performance on reasoning over unseen knowledge facts, it results in catastrophically forgetting the previously learnt knowledge. Our main results suggest that PLMs cannot yet perform reliable deductive reasoning, demonstrating the importance of controlled examinations and probing of PLMs' deductive reasoning abilities; we reach beyond (misleading) task performance, revealing that PLMs are still far from robust reasoning capabilities, even for simple deductive tasks.

Data Augmentation and Learned Layer Aggregation for Improved Multilingual Language Understanding in Dialogue

“…Further, one issue which might arise from further specialising a model to a given task/dataset is a well-known phenomenon of catastrophic forgetting: pretrained language models are prone to forgetting previously learnt knowledge or skills when tuned on new data (De Cao et al, 2021;Yuan et al, 2023). To evaluate whether the models would retain their ability to respond faithfully to examples considerably different from ones seen during fine-tuning, we evaluate the models tuned on Faith-Dial on TopiOCQA.…”