On language models’ cognitive biases in reading time prediction

Patrick Haller, Lena Sophia Bolliger, Lena Ann Jäger

Keywords: cognitive interpretability, individual differences, reading time analysis, language modeling

TL;DR: We investigate what type of psycholinguistic subjects language models emulate by assessing the predictive power of surprisal and entropy effects modulated by individual psychometric scores (e.g., reading fluency).

Abstract:

To date, most investigations on surprisal and entropy effects in reading have been conducted on the group-level, disregarding individual differences. In this work, we revisit the predictive power (PP) of different language models' (LMs') surprisal and entropy measures on data of human reading times by incorporating information of language users' cognitive capacities. To do so, we assess the PP of surprisal and entropy estimated from generative LMs on reading data from subjects for which scores from psychometric tests targeting different cognitive domains are available.

Specifically, we investigate if modulating surprisal and entropy relative to the readers' cognitive scores increases prediction accuracy of reading times, and we examine whether LMs exhibit systematic biases in the prediction of reading times for cognitively high- or low-scoring groups, allowing us to investigate what type of psycholinguistic subjects a given LM emulates.

We find that incorporating cognitive capacities mostly increases PP of surprisal and entropy on reading times, and that individuals performing high in cognitive tests are less sensitive to predictability effects. Our results further suggest that the analyzed LMs emulate readers with lower verbal intelligence, suggesting that for a given target group (i.e., individuals with high verbal intelligence), these LMs provide less accurate predictability estimates. Finally, our study underlines the value of incorporating individual-level information to gain insights into how LMs operate internally.

Submission Number: 59