Language Models Explain Word Reading Times Better Than Empirical Predictability.

Area of Science:

• Cognitive psychology
• Computational linguistics
• Neuroscience of language

Background:

• Visual-orthographic access to the mental lexicon is influenced by word length and frequency.
• Syntactic and semantic factors in reading are traditionally captured by cloze completion probability (CCP).
• Recent research suggests probabilistic language models offer superior explanations for these effects.

Purpose of the Study:

• To compare the predictive power of cloze completion probability (CCP) with three probabilistic language models (n-gram, topic, RNN) for reading times.
• To investigate how different language models capture syntactic and semantic influences on lexical access.
• To determine which models best predict various eye-movement measures during reading.

Main Methods:

• Compared CCP with symbolic n-gram, topic, and recurrent neural network (RNN) models.
• Used models to predict word viewing times (single fixation duration, gaze duration, total viewing time) in English and German eye-tracking data.
• Employed linear and non-linear analyses (generalized additive models) to assess prediction accuracy.

Main Results:

• All probabilistic language models showed higher correlations with eye-movement measures than CCP.
• N-gram and RNN models consistently predicted reading performance better than topic models or CCP.
• N-gram models excelled at predicting short-range access effects, while RNNs better predicted early preprocessing of subsequent words.

Conclusions:

• Probabilistic language models provide more nuanced explanations for syntactic and semantic effects in reading than CCP.
• Different language models capture distinct cognitive processes involved in lexical consolidation during reading.
• These models serve as valuable, algorithmically defined blueprints for understanding human reading mechanisms.