On the Systematicity of Probing Contextualized Word Representations: The Case of Hypernymy in BERT

Proceedings of the Ninth Joint Conference on Lexical and Computational Semantics (*SEM), pages 88–102 Barcelona, Spain (Online), December 12–13, 2020 88 On the Systematicity of Probing Contextualized Word Representations: The Case of Hypernymy in BERT Abhilasha Ravichander♦* , Eduard Hovy♦, Kaheer Suleman♥ Adam Trischler♥♠, Jackie Chi Kit Cheung♠ ♦Carnegie Mellon University, Pittsburgh, PA ♥Microsoft Research, Montreal, Canada ♠McGill University, Montreal, Canada {aravicha, ehovy}@cs.cmu.edu {adam.trischler, kasulema}@microsoft.com {jcheung}@cs.mcgill.ca Abstract Contextualized word representations have be- come a driving force in NLP, motivating widespread interest in understanding their ca- pabilities and the mechanisms by which they operate. Particularly intriguing is their abil- ity to identify and encode conceptual abstrac- tions. Past work has probed BERT representa- tions (Devlin et al., 2019) for this competence, finding that BERT can correctly retrieve noun hypernyms in cloze tasks. In this work, we ask the question: do probing studies shed light on systematic knowledge in BERT representa- tions? As a case study, we examine hyper- nymy knowledge encoded in BERT represen- tations. In particular, we demonstrate through a simple consistency probe that the ability to correctly retrieve hypernyms in cloze tasks, as used in prior work, does not correspond to sys- tematic knowledge in BERT. Our main conclu- sion is cautionary: even if BERT demonstrates high probing accuracy for a particular compe- tence, it does not necessarily follow that BERT ‘understands’ a concept, and it cannot be ex- pected to systematically generalize across ap- plicable contexts.1 1 Introduction Hierarchical representations of concepts play a cen- tral role in reasoning and understanding natural language (Wellman and Gelman, 1992). They have long been studied as a core NLP objective in their own right through tasks requiring the identification of hypernyms (Hearst, 1992; Snow et al., 2005, 2006), and as components for use in downstream * Part of this work was done during an internship at Microsoft Research. 1Diagnostic framework available at //github.com/AbhilashaRavichander/https: probe-generalization. This work is licensed under a Creative Commons Attri- bution 4.0 International License. License details: http: //creativecommons.org/licenses/by/4.0/. A car is a _____ cars are _____ vehicle trees Singular Grammatical Number Probe Plural Grammatical Number Probe Figure 1: Illustration of BERT’s inconsistent predic- tions on singular and plural hypernymy probes. NLP tasks, such as recognizing textual entailment (RTE), metaphor detection, text generation and question answering (QA) (Girju et al., 2003; Dagan et al., 2006; Prager et al., 2008; Mirkin et al., 2009; Akhmatova and Dras, 2009; Mohler et al., 2013; Biran and McKeown, 2013; Yahya et al., 2013). Recently, Pretrained Language Models (PLMs), such as BERT (Devlin et al., 2019), have emerged as a popular and successful approach to a variety of NLP tasks. Thus, there has been community interest in evaluating their representations for the ‘knowledge’ they contain, including information about concept abstraction (Ettinger, 2020; Talmor et al., 2019; Jiang et al., 2020; Petroni et al., 2019). We distinguish research that investigates knowl- edge encoded in BERT through two broad perspec- tives: instrumentative and agentive. We view the instrumentative perspective as treating PLMs as a tool to …

Content truncated to 500 words