Measuring and Improving Consistency in Pretrained Language Models

Measuring and Improving Consistency in Pretrained Language Models Yanai Elazar1,2 Nora Kassner3 Shauli Ravfogel1,2 Abhilasha Ravichander4 Eduard Hovy4 Hinrich Sch¨utze3 Yoav Goldberg1,2 1Computer Science Department, Bar Ilan University, Israel 2Allen Institute for Artificial Intelligence, United States 3Center for Information and Language Processing (CIS), LMU Munich, Germany 4Language Technologies Institute, Carnegie Mellon University, United States {yanaiela,shauli.ravfogel,yoav.goldberg}@gmail.com kassner@cis.lmu.de {aravicha,hovy}@cs.cmu.edu Abstract Consistency of a model—that is, the invari- ance 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 re- spect to factual knowledge? To this end, we create PARAREL , a high-quality resource of cloze-style query English paraphrases. It con- tains 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 represent- ing knowledge robustly. Finally, we propose a method for improving model consistency and experimentally demonstrate its effectiveness.1 1 Introduction Pretrained Language Models (PLMs) are large neural networks that are used in a wide variety of NLP tasks. They operate under a pretrain-finetune paradigm: Models are first pretrained over a large text corpus and then finetuned on a downstream task. PLMs are thought of as good language en- coders, supplying basic language understanding capabilities that can be used with ease for many downstream tasks. A desirable property of a good language un- derstanding model is consistency: the ability to make consistent decisions in semantically equiv- alent contexts, reflecting a systematic ability to generalize in the face of language variability. 1The code and resource are available at: https:// github.com/yanaiela/pararel. Examples of consistency include: predicting the same answer in question answering and read- ing comprehension tasks regardless of paraphrase (Asai and Hajishirzi, 2020); making consistent assignments in coreference resolution (Denis and Baldridge, 2009; Chang et al., 2011); or making summaries factually consistent with the original document (Kryscinski et al., 2020). While consis- tency is important in many tasks, nothing in the training process explicitly targets it. One could hope that the unsupervised training signal from large corpora made available to PLMs such as BERT or RoBERTa (Devlin et al., 2019; Liu et al., 2019) is sufficient to induce consistency and transfer it to downstream tasks. In this paper, we show that this is not the case. The recent rise of PLMs has sparked a discus- sion about whether these models can be used as Knowledge Bases (KBs) (Petroni et al., 2019; 2020; Davison et al., 2019; Peters et al., 2019; Jiang et al., 2020; Roberts et al., 2020). Consis- tency is a key property of KBs and is particularly important for automatically constructed KBs. One of the biggest appeals of using a PLM as a KB is that we can query it in natural language— instead of relying on a specific …

Content truncated to 500 words