Does BERT really agree ? Fine-grained Analysis of Lexical Dependence on a Syntactic Task

Findings of the Association for Computational Linguistics: ACL 2022, pages 2309 - 2315 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Does BERT really agree ? Fine-grained Analysis of Lexical Dependence on a Syntactic Task Karim Lasri@,B Alessandro LenciB Thierry Poibeau@ @Lattice (École Normale Supérieure-PSL, CNRS, U. Sorbonne Nouvelle) BUniversity of Pisa karim.lasri@ens.psl.eu alessandro.lenci@unipi.it thierry.poibeau@ens.psl.eu Abstract Although transformer-based Neural Language Models demonstrate impressive performance on a variety of tasks, their generalization abil- ities are not well understood. They have been shown to perform strongly on subject-verb number agreement in a wide array of settings, suggesting that they learned to track syntactic dependencies during their training even with- out explicit supervision. In this paper, we ex- amine the extent to which BERT is able to per- form lexically-independent subject-verb num- ber agreement (NA) on targeted syntactic tem- plates. To do so, we disrupt the lexical patterns found in naturally occurring stimuli for each targeted structure in a novel fine-grained anal- ysis of BERT’s behavior. Our results on nonce sentences suggest that the model generalizes well for simple templates, but fails to perform lexically-independent syntactic generalization when as little as one attractor is present. 1 Introduction Every English speaker would judge as grammatical the sentences in (1a)-(1b), but not those in (1c)- (1d), despite that they are all meaningless: (1) a. Colourless green ideas sleep furiously. b. Colourless green ideas that cook the door sleep furiously. c. *Colourless green ideas sleeps furiously. d. *Colourless green ideas that cook the door sleeps furiously. At least since Chomsky (1957), data like this has been taken as evidence that natural language gram- mars contain abstract syntactic rules that (i) are independent of the meaning of lexical items and (ii) obey hierarchical, rather than linear constraints. Number agreement (henceforth NA) between the subject (the cue) and the verb (the target) of the same clause in English is one of such rules (Corbett, 2003). In fact, (1d) is ungrammatical, even though the closest noun door (typically referred to as at- tractor) has the same number as sleeps, because the noun belongs to an embedded relative clause. These NA properties have made it one of the pre- ferred test beds to investigate the ability of neural language models (NLMs) to learn abstract, hier- archical syntactic structures (Linzen et al., 2016; Gulordava et al., 2018; Marvin and Linzen, 2018; Goldberg, 2019; Bacon and Regier, 2019; Lakretz et al., 2019). Although recurrent and transformer- based NLMs have been shown to possess syntactic abilities on the task, their nature is not fully under- stood (Baroni, 2019). Can NLMs really perform lexically-independent number agreement, regardless of the syntactic struc- ture? To answer this question, we test BERT (De- vlin et al., 2019) against the NA task while control- ling both the syntactic constructions and the mean- ingfulness of the stimuli presented to the model. Our experiments provide two main findings. Contrarily to previous observations that BERT per- forms fairly well on Gulordava et al.’s (2018) syn- tactically well-formed but meaningless sentences (Goldberg, …

Content truncated to 500 words