Measuring and Improving Consistency in Pretrained Language Models

TL;DR

This paper evaluates the consistency of pretrained language models with respect to factual knowledge, introduces a new paraphrase dataset, and proposes a method to enhance model consistency.

Contribution

It introduces ParaRel, a high-quality paraphrase dataset for evaluating consistency, and presents a novel method to improve the consistency of pretrained language models.

Findings

PLMs show poor consistency with high variance across relations.

Representational spaces of PLMs are poorly structured for robust knowledge representation.

The proposed method effectively improves model consistency.

Abstract

Consistency of a model -- that is, the invariance of its behavior under meaning-preserving alternations in its input -- is a highly desirable property in natural language processing. In this paper we study the question: Are Pretrained Language Models (PLMs) consistent with respect to factual knowledge? To this end, we create ParaRel, a high-quality resource of cloze-style query English paraphrases. It contains a total of 328 paraphrases for 38 relations. Using ParaRel, we show that the consistency of all PLMs we experiment with is poor -- though with high variance between relations. Our analysis of the representational spaces of PLMs suggests that they have a poor structure and are currently not suitable for representing knowledge robustly. Finally, we propose a method for improving model consistency and experimentally demonstrate its effectiveness.