Unsupervised Extractive Summarization of Psychedelic User Experience Reports

Abstract

A bstract Contemporary psychedelic research highlights the value of user experience reports, yet their verbose, subjective nature poses challenges for clinical utility. This is the first study to pioneer unsupervised automatic text summarization of psychedelic user experience reports, a domain where no human-annotated reference summaries exist. To address this gap, we developed a custom scoring function that integrates semantic coverage, narrative coherence, and a novel experiential preservation metric, enabling effective model training and hyperparameter tuning. We utilized three established extractive methods: LexRank, LSA with HDBSCAN clustering, and SBERT with Maximal Marginal Relevance, on 1,200 reports involving LSD, psilocybin, and DMT. Using GPT-4 as a calibrated rater under a structured rubric, supplemented by TOPSIS aggregation, results showed LexRank achieving the highest overall balance with SBERT excelling in content coverage and experiential depth but lagging in coherence. Our findings revealed trade-offs between content richness and narrative fluency, with performance varying across substance types due to differences in narrative structure and phenomenology. Limitations included reliance on extractive methods, lack of reference data, and sensitivity to scoring design. Future work should extend to abstractive methods, alternative weighting schemes, and expert adjudication to develop clinically usable summarization systems for psychedelic science.