Investigating BERT’s Knowledge of Language: Five Analysis Methods with NPIs

Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing, pages 2877–2887, Hong Kong, China, November 3–7, 2019. c⃝2019 Association for Computational Linguistics 2877 Investigating BERT’s Knowledge of Language: Five Analysis Methods with NPIs Alex Warstadt,†,1,2 Yu Cao,†,3 Ioana Grosu,†,2 Wei Peng,†,3 Hagen Blix,†,1 Yining Nie,†,1,2 Anna Alsop,†,2 Shikha Bordia,†,3 Haokun Liu,†,3 Alicia Parrish,†,2,3 Sheng-Fu Wang,†,3 Jason Phang,†,1,3 Anhad Mohananey,†,1,3 Phu Mon Htut,†,3 Paloma Jeretiˇc,†,1,2 and Samuel R. Bowman New York University †Equal contribution with roles given below; order assigned randomly. Correspondence: bowman@nyu.edu 1Framing and organizing the paper 2Creating diagnostic data 3Constructing and running experiments Abstract Though state-of-the-art sentence representa- tion models can perform tasks requiring sig- nificant knowledge of grammar, it is an open question how best to evaluate their grammat- ical knowledge. We explore five experimen- tal methods inspired by prior work evaluat- ing pretrained sentence representation models. We use a single linguistic phenomenon, nega- tive polarity item (NPI) licensing in English, as a case study for our experiments. NPIs like any are grammatical only if they appear in a licensing environment like negation (Sue doesn’t have any cats vs. *Sue has any cats). This phenomenon is challenging because of the variety of NPI licensing environments that exist. We introduce an artificially generated dataset that manipulates key features of NPI licensing for the experiments. We find that BERT has significant knowledge of these fea- tures, but its success varies widely across dif- ferent experimental methods. We conclude that a variety of methods is necessary to reveal all relevant aspects of a model’s grammatical knowledge in a given domain. 1 Introduction Recent sentence representation models have at- tained state-of-the-art results on language under- standing tasks, but standard methodology for eval- uating their knowledge of grammar has been slower to emerge. Recent work evaluating gram- matical knowledge of sentence encoders like BERT (Devlin et al., 2018) has employed a vari- ety of methods. For example, Shi et al. (2016), Ettinger et al. (2016), and Tenney et al. (2019) use probing tasks to target a model’s knowledge of particular grammatical features. Marvin and Linzen (2018) and Wilcox et al. (2019) compare language models’ probabilities for pairs of mini- mally different sentences differing in grammatical acceptability. Linzen et al. (2016), Warstadt et al. (2018), and Kann et al. (2019) use Boolean accept- ability judgments inspired by methodologies in generative linguistics. However, we have not yet seen any substantial direct comparison between these methods, and it is not yet clear whether they tend to yield similar conclusions about what a given model knows. We aim to better understand the trade-offs in task choice by comparing different methods in- spired by previous work to evaluate sentence un- derstanding models in a single empirical domain. We choose as our case study negative polarity item (NPI) licensing in English, an empirically rich phenomenon widely discussed in the theoret- ical linguistics literature (Kadmon and Landman, 1993; Giannakidou, 1998; Chierchia, 2013, a.o.). NPIs are words or …

Content truncated to 500 words