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

Abstract

We present a targeted, scaled-up comparison of incremental processing in humans and neural language models by collecting by-word reaction time data for sixteen different syntactic test suites across a range of structural phenomena. Human reaction time data comes from a novel online experimental paradigm called the Interpolated Maze task. We compare human reaction times to by-word probabilities for four contemporary language models, 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 human and model 'accuracy' scores (à la Marvin and Linzen (2018)) about equal. However, 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 sentences. Specifically, when models encounter syntactic violations they fail to accurately predict the longer reaction times observed in the human data. These results call into question whether contemporary language models are approaching human-like performance for sensitivity to syntactic violations.1. S d (was impressed) < Sc(was impressed) 2. S d (was impressed) < S b (was impressed)

Incremental Processing of Principle B: Mismatches Between Neural Models and Humans

“…The present study argues that autoregressive models do not (uniformly) process pronouns like humans. We showed that models fail to capture the qualitative patterns of human incremental coreference processing, in addition to underestimating processing costs in constructions already noted in the literature (see van Schijndel and Linzen, 2021;Wilcox et al, 2021b). Models appear to learn only aspects of Principle B that have predictable reflexes in training data.…”

“…Recent work has placed increased attention on finer-grained comparisons between neural models and humans (e.g., van Schijndel and Linzen, 2021;Wilcox et al, 2021b;Paape and Vasishth, 2022). The growing consensus is that neural models underestimate the processing costs seen with humans, while nonetheless capturing the broad patterns (see Wilcox et al, 2021b).…”

Testing the limits of natural language models for predicting human language judgments

“…Unlike convolutional neural networks, whose architectural design principles are roughly inspired by biological vision [Lindsay, 2021], the design of current neural network language models is largely uninformed by psycholinguistics and neuroscience. And yet, there is an ongoing effort to adopt and adapt neural network language models to serve as computational hypotheses of how humans process language, making use of a variety of different architectures, training corpora, and training tasks [e.g., Wehbe et al, 2014, Toneva and Wehbe, 2019, Heilbron et al, 2020, Jain et al, 2020, Lyu et al, 2021, Schrimpf et al, 2021, Wilcox et al, 2021, Goldstein et al, 2022, Caucheteux and King, 2022. We found that recurrent neural networks make markedly human-inconsistent predictions once pitted against transformer-based neural networks.…”

A-maze of Natural Stories: Comprehension and surprisal in the Maze task

“…Another recent variant of the Maze task is interpolated I-maze, which uses a mix of real word distractors (generated via the A-maze process) and non-word distractors (Vani, Wilcox, and Levy 2021;Wilcox, Vani, and Levy 2021). The presence of real word distractors encourages close attention to the sentential context, while non-words can be used as distractors where the word in the sentence is itself ungrammatical or highly unexpected, and/or it is important that the predictability of the distractor in the context is perfectly well-balanced (at zero) across all experimental conditions.…”