Word Frequency Does Not Predict Grammatical Knowledge in Language Models

Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, pages 4040–4054, November 16–20, 2020. c⃝2020 Association for Computational Linguistics 4040 Word Frequency Does Not Predict Grammatical Knowledge in Language Models Charles Yu, Ryan Sie, Nico Tedeschi, and Leon Bergen {cty002,rdsie,njtedeschi,lbergen}@ucsd.edu University of California, San Diego Abstract Neural language models learn, to varying de- grees of accuracy, the grammatical properties of natural languages. In this work, we in- vestigate whether there are systematic sources of variation in the language models’ accuracy. Focusing on subject-verb agreement and re- flexive anaphora, we find that certain nouns are systematically understood better than oth- ers, an effect which is robust across grammati- cal tasks and different language models. Sur- prisingly, we find that across four orders of magnitude, corpus frequency is unrelated to a noun’s performance on grammatical tasks. Fi- nally, we find that a novel noun’s grammatical properties can be few-shot learned from vari- ous types of training data. The results present a paradox: there should be less variation in grammatical performance than is actually ob- served. 1 Introduction Neural language models (Howard and Ruder, 2018; Devlin et al., 2019; Dai et al., 2019; Yang et al., 2019; Radford et al., 2019) have achieved success in both text prediction and downstream tasks such as question-answering, text classification, and natu- ral language inference. The strong performance of these models raises scientific questions about the knowledge they have acquired, in particular, about the abstractness and generality of their linguistic representations. Previous work has investigated the linguistic rep- resentations of neural language models in several domains, and found varying evidence for how lin- guistically adequate these representations are (Lau et al., 2017; Marvin and Linzen, 2018; Goldberg, 2019; Futrell et al., 2019). This work has em- ployed psycholinguistic methodology in order to elicit grammatical judgments from these models, inferring the models’ underlying representations from the patterns of judgments. In the current work, we focus on the variation in grammatical knowledge that potentially exists within a neural language model. Just as in human psycholinguistic tasks, previous work on neural LMs has observed variability in grammatical judg- ments between different sentences; not all viola- tions of a grammatical constraint are judged to be equally bad. It is not clear, however, whether there are systematic sources of variation in these judg- ments, and if so, what the sources are. We will focus on variation among lexical items, using English subject-verb agreement and reflexive anaphora as a case study. We first ask whether lan- guage models learn the grammatical properties of some nouns more accurately than for others. We do this by measuring the accuracy of language models when making grammatical judgments involving dif- ferent nouns. We find systematic variation among nouns: nouns that perform well on one task or lan- guage model are more likely to perform well on other tasks or other language models. We then consider possible sources of the observed varia- tion between nouns, finding that the grammatical properties of nouns are paradoxically easy …

Content truncated to 500 words