In the study of language production, Cloze and trigrams are among the most established methods to quantify the predictability of speech. Both of these methods have been extensively used to study language production in healthy subjects and people with neurocognitive disorders. However, both of these methods have limitations that impair their potential to be used as objective measures of psychosis. ... read moreIn this study, we introduce the use of a state-of-the-art machine-learning language model, GPT-2, as an alternative measure of speech predictability. By comparing GPT-2’s probability calculations to Cloze and trigram probabilities for the same sample data, we found that GPT-2 can highly correlate to both measures. We then focused on the potential application of GPT-2 probability calculations to study speech production in schizophrenia, suggesting ways in which it can be used in place of n-grams and Cloze, and exemplifying how it can be used to test hypotheses on the underlying working memory deficits that have been linked to schizophrenia.read less
