A Targeted Assessment of Incremental Processing in Neural Language Models and Humans

Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing, pages 939–952 August 1–6, 2021. ©2021 Association for Computational Linguistics 939 A Targeted Assessment of Incremental Processing in Neural Language Models and Humans Ethan Gotlieb Wilcox Harvard University Department of Linguistics wilcoxeg@g.harvard.edu Pranali Vani MIT Brain and Cognitive Science pvani@mit.edu Roger P. Levy MIT Brain and Cognitive Science rplevy@mit.edu Abstract We present a targeted, scaled-up comparison of incremental processing in humans and neu- ral language models by collecting by-word re- action time data for sixteen different syntac- tic test suites across a range of structural phe- nomena. Human reaction time data comes from a novel online experimental paradigm called the Interpolated Maze task. We com- pare human reaction times to by-word proba- bilities for four contemporary language mod- els, with different architectures and trained on a range of data set sizes. We find that across many phenomena, both humans and language models show increased processing difficulty in ungrammatical sentence regions with hu- man and model ‘accuracy’ scores (`a la Mar- vin and Linzen (2018)) about equal. How- ever, although language model outputs match humans in direction, we show that models systematically under-predict the difference in magnitude of incremental processing difficulty between grammatical and ungrammatical sen- tences. Specifically, when models encounter syntactic violations they fail to accurately pre- dict the longer reaction times observed in the human data. These results call into question whether contemporary language models are approaching human-like performance for sen- sitivity to syntactic violations. 1 Introduction A substantial body of work has investigated con- temporary language models (LMs) by assessing whether their behavior is consistent with the rules of syntax (Hu et al., 2020; Marvin and Linzen, 2018; Warstadt et al., 2020).1 Among other structures, these studies have investigated agree- ment (Linzen et al., 2016; Gulordava et al., 2018) 1Data and code for this paper can be found online at https://github.com/wilcoxeg/ targeted-assessment-imaze long distance dependencies (Wilcox et al., 2018), pronominal and particle licensing (Jumelet and Hupkes, 2018; Futrell et al., 2019), and expecta- tions for phrase-level constituents (Futrell et al., 2018). Many of the studies which report aggre- gate behavior across a broad number of phenom- ena focus on accuracy scores, or the proportion of time LMs or human subjects in an online ex- periment prefer a grammatical variant in match- ing grammatical / ungrammatical sentence pairs. While these investigations provide much insight, they collapse a crucial dimension of comparison, namely the difference in magnitude between the grammatical and ungrammatical conditions. As long as the direction of their predictions are the same, an LM which finds grammatical conditions only marginally worse than their corresponding ungrammatical counterpart will receive the same score as a model that displays large differences between the two conditions. At the same time, a related line of work has investigated the quantitative relationship between incremental predictions of language models and human reaction times (Hale, 2001; Levy, 2008). Smith and Levy (2013) found that …

Content truncated to 500 words