Multilingual Language Models Predict Human Reading Behavior

TL;DR

This paper investigates whether large multilingual transformer models can accurately predict human reading behavior across multiple languages, revealing their implicit encoding of language importance similar to human processing.

Contribution

It demonstrates that multilingual transformer models like BERT and XLM can effectively predict human reading patterns across different languages, highlighting their potential to model human sentence processing.

Findings

BERT and XLM models predict eye tracking features accurately

Models show cross-language generalization in reading behavior prediction

Transformer models encode language importance akin to human cognition

Abstract

We analyze if large language models are able to predict patterns of human reading behavior. We compare the performance of language-specific and multilingual pretrained transformer models to predict reading time measures reflecting natural human sentence processing on Dutch, English, German, and Russian texts. This results in accurate models of human reading behavior, which indicates that transformer models implicitly encode relative importance in language in a way that is comparable to human processing mechanisms. We find that BERT and XLM models successfully predict a range of eye tracking features. In a series of experiments, we analyze the cross-domain and cross-language abilities of these models and show how they reflect human sentence processing.