Morphosyntactic but not lexical corpus-based probabilities can substitute for cloze probabilities in reading...

Area of Science:

• Psycholinguistics
• Computational Linguistics
• Cognitive Science

Background:

• Predicting upcoming words during language processing is crucial for comprehension.
• Traditional methods like the cloze task rely on human judgment to estimate predictability.
• Corpus-based language models offer a potential alternative for quantifying predictability.

Purpose of the Study:

• To investigate a corpus-based approach for estimating lexical and morphosyntactic predictability.
• To compare the predictive power of corpus-based and human-based (cloze) predictability measures on reading times.
• To determine if corpus-based measures can substitute for cloze probabilities in psycholinguistic experiments.

Main Methods:

• Collected cloze and corpus-based probabilities for words in 144 Russian sentences.
• Analyzed eye movement data during reading to measure reading times.
• Correlated predictability measures and assessed their variance explained in reading times.

Main Results:

• Strong correlation found between cloze and corpus-based predictability measures.
• Both measures explained similar variance in reading times for lexical and morphosyntactic predictability.
• Corpus-based morphosyntactic probabilities can substitute for cloze probabilities, unlike lexical ones.

Conclusions:

• Corpus-based models are viable alternatives to cloze tasks for estimating morphosyntactic predictability in reading research.
• In morphologically rich languages like Russian, predicting isolated morphosyntactic features is more successful than predicting full word forms.