Uncovering Constraint-Based Behavior in Neural Models via Targeted Fine-Tuning

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 1159–1171 August 1–6, 2021. ©2021 Association for Computational Linguistics 1159 Uncovering Constraint-Based Behavior in Neural Models via Targeted Fine-Tuning Forrest Davis and Marten van Schijndel Department of Linguistics Cornell University {fd252|mv443}@cornell.edu Abstract A growing body of literature has focused on detailing the linguistic knowledge embedded in large, pretrained language models. Existing work has shown that non-linguistic biases in models can drive model behavior away from linguistic generalizations. We hypothesized that competing linguistic processes within a language, rather than just non-linguistic model biases, could obscure underlying linguistic knowledge. We tested this claim by exploring a single phenomenon in four languages: En- glish, Chinese, Spanish, and Italian. While human behavior has been found to be similar across languages, we find cross-linguistic vari- ation in model behavior. We show that compet- ing processes in a language act as constraints on model behavior and demonstrate that tar- geted fine-tuning can re-weight the learned constraints, uncovering otherwise dormant lin- guistic knowledge in models. Our results sug- gest that models need to learn both the linguis- tic constraints in a language and their relative ranking, with mismatches in either producing non-human-like behavior. 1 Introduction Ever larger pretrained language models continue to demonstrate success on a variety of NLP bench- marks (e.g., Devlin et al., 2019; Brown et al., 2020). One common approach for understanding why these models are successful is centered on infer- ring what linguistic knowledge such models ac- quire (e.g., Linzen et al., 2016; Hewitt and Man- ning, 2019; Hu et al., 2020; Warstadt et al., 2020a). Linguistic knowledge alone, of course, does not fully account for model behavior; non-linguistic heuristics have also been shown to drive model behavior (e.g., sentence length; see McCoy et al., 2019; Warstadt et al., 2020b). Nevertheless, when looking across a variety of experimental methods, models appear to acquire some grammatical knowl- edge (see Warstadt et al., 2019). However, investigations of linguistic knowledge in language models are limited by the overwhelm- ing prominence of work solely on English (though see Gulordava et al., 2018; Ravfogel et al., 2018; Mueller et al., 2020). Prior work has shown non- linguistic biases of neural language models mimic English-like linguistic structure, limiting the gener- alizability of claims founded on English data (e.g., Dyer et al., 2019; Davis and van Schijndel, 2020b). In the present study, we show via cross-linguistic comparison, that knowledge of competing linguis- tic constraints can obscure underlying linguistic knowledge. Our investigation is centered on a single dis- course phenomena, implicit causality (IC) verbs, in four languages: English, Chinese, Spanish, and Italian. When an IC verb occurs in a sentence, interpretations of pronouns are affected: (1) a. Lavender frightened Kate because she was so terrifying. b. Lavender admired Kate because she was so amazing. In (1), both Lavender and Kate agree in gender with she, so both are possible antecedents. How- ever, English speakers overwhelmingly interpret she as …

Content truncated to 500 words