Dream reports are easier for large language models to predict than Wikipedia articles, as measured by lower perplexity scores, indicating that dream content is less 'surprising' to these models than general web text. The models also detected differences in dream reports based on gender, visual impairment, and clinical status, mirroring patterns found in prior research. This suggests that machine learning tools can effectively model dream narratives and may capture subtle group-level variations.
Dream research usually depends on human experts manually scoring dream reports, a time-consuming process. While natural language processing tools have been explored for automatic analysis, they could not reason over a report's full context, needed extensive preprocessing, and were rarely validated against manual scoring. This work used large language models, both off-the-shelf and bespoke, to replicate manual annotation of dream reports, focusing on emotions. The off-the-shelf method performed poorly, likely due to linguistic differences across individuals. In contrast, the bespoke text classification method achieved high performance and was robust against biases. This approach may enable analysis of large dream datasets and improve reproducibility and comparability across studies.