[Paper Review] Measuring and Improving Consistency in Pretrained Language Models

Yanai Elazar, Nora Kassner+5 more|arXiv (Cornell University)|Feb 1, 2021

Topic Modeling77 references32 citations

TL;DR

The paper introduces ParaRel to measure consistency in pretrained language models (PLMs) via paraphrase cloze patterns, finds widespread inconsistency across models, and proposes a KL-divergence based consistency loss to improve it.

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.

Motivation & Objective

• Assess whether PLMs maintain consistent factual predictions under meaning-preserving paraphrases.
• Introduce ParaRel, a high-quality paraphrase-based benchmark for 38 relations with 328 patterns.
• Evaluate knowledge extraction and consistency across BERT, RoBERTa, and ALBERT.
• Propose and validate a consistency loss to improve model consistency on unseen relations.

Proposed method

• Define a consistency framework using data triples D_i and paraphrase patterns P_i for relation r_i with restricted candidate sets to KB objects.
• Construct ParaRel by augmenting base patterns, incorporating LPAQA paraphrases, SPIKE-derived patterns, and expert curation; ensure high inter-annotator agreement.
• Evaluate consistency (Consistency), knowledge (Accuracy), and joint metric (Consistent-Acc) across multiple PLMs.
• Introduce a consistency loss based on two-sided KL divergence between pattern-specific output distributions, combined with MLM loss, and train with relation-wise batching.

Experimental results

Research questions

• RQ1 Is PLM behavior invariant to paraphrase-like variations in factual queries?
• RQ2 How well does ParaRel capture consistency across a diverse set of relations and patterns?
• RQ3 Can a dedicated consistency loss improve cross-pattern agreement on unseen relations?

Key findings

• PLMs show overall poor consistency with high variance by relation.
• BERT-based models generally achieve higher consistency and accuracy than RoBERTa and ALBERT in this setup.
• Consistency and accuracy are positively correlated across models, but larger training data does not always yield better consistency (an exception exists).
• Syntax robustness exists but knowledge extraction remains sensitive to surface variations, indicating surface form effects beyond syntax.
• Applying the proposed consistency loss improves consistency on unseen relations (evidence from the study).