Language Models Learn Sentiment and Substance from 11,000 Psychoactive Experiences

The paper situates itself in the challenge of quantifying the rich, temporally unfolding subjective states induced by psychoactive substances, noting that these subjective qualities correlate with clinical outcomes in psychedelic-assisted therapy. The authors argue conventional questionnaires compress a vast experiential space into a few researcher-selected scalar measures and therefore miss temporal dynamics (for example peak-end and primacy effects) that may condition therapeutic benefit. They propose that natural language contained in patient or user testimonials, combined with modern natural language processing (NLP), can provide multi-dimensional, temporally sensitive measures of subjective experience and thereby help link phenomenology to neurobiology. G. and colleagues set out to build and compare three machine learning approaches that together produce neurobiologically informed, temporally resolved portraits of drug-induced experiences. Using a large corpus of 11,816 Erowid testimonials, pretrained transformer encoders fine-tuned in supervised and transfer-learning paradigms were used to recover multi-dimensional sentiment trajectories and other metadata, while canonical correlation analysis (CCA) linked testimonial word-usage to receptor affinity fingerprints for 52 drugs. The stated aim is to demonstrate that these complementary methods converge on meaningful experiential structure that can be anchored to pharmacology and mapped to cortical gene-expression, thereby suggesting routes toward data-driven neurofeedback and personalised therapeutic guidance.

Data sources comprised 11,816 Erowid user testimonials, receptor affinity data at up to 61 receptor subtypes for 52 drugs (primarily from the Psychoactive Drug Screening Program, supplemented by Rickli et al. for eight phenethylamines), RNA gene-expression measures for 200 brain regions from the Allen Brain Atlas, and 58,000 Reddit posts annotated with 28 emotions (the GoEmotions dataset). Chemical and pharmacologic class labels were taken from Psychonaut Wiki. Text preprocessing removed occurrences of drug names (scientific, common, colloquial and misspellings) to avoid trivial cues, tokenised the texts and capped model inputs at 512 tokens. A sliding-window inference scheme was applied to longer testimonials to build trajectories from contiguous blocks; windows smaller than the testimonial were split and each block processed, while windows larger than the text were padded. Dynamic Time Warping (DTW) was used to compare averaged sentiment trajectories between drugs and classes. Two transformer-based models were fine-tuned. Both used a base bidirectional transformer (BERT) encoder with approximately 109 million parameters; pooled outputs fed through dropout and task-specific dense heads. BERTowid is a multitask, multi-label supervised model trained directly on Erowid excerpts to predict categorical labels (drug identity, chemical and pharmacologic classes, metadata tags, self-reported gender) and regression targets (age, receptor affinities, and CCA component weights). Classification losses minimised cross-entropy, regressions minimised mean squared error; due to sensitivity when optimising mixed losses, the authors trained and serialized separate multitask models for classification and regression. Hyperparameters included AdamW optimisation, learning rate 1e-5, batch size 32 and 16 epochs; a typical run with window size 64 words required ~3 hours on an NVIDIA V-100 GPU. Imbalanced labels were explored with weighted cross-entropy but with mixed effects on precision. BERTiment is a transfer-learning model: the same base BERT encoder was fine-tuned on the GoEmotions Reddit corpus to predict 28 binary emotion labels (training split 70-20-10). Hyperparameters matched BERTowid except dropout was set to 0.5; training ran for 16 epochs (~2 hours on a V-100). BERTiment produced temporally resolved emotion predictions when applied to testimonial window series. Canonical Correlation Analysis (CCA) provided an independent, linear linkage between bag-of-words representations of whole testimonials and receptor-affinity fingerprints. The scikit-learn CCA implementation was used to extract paired components (word weightings and receptor weightings). Receptor-component weights were projected to cortex using Allen Brain Atlas RNA expression quantities. The authors contrasted the CCA (linear, bag-of-words, whole-text) with the transformer models (nonlinear, positionally encoded, excerpt-based) and compared cross-model consistency.

The analysis used 11,816 Erowid testimonials and produced convergent results across three analytical approaches. BERTowid achieved substantive discrimination across multiple pharmacological granularities: 52 drugs, 22 chemical types, 10 pharmacologic classes, 30 receptor subtypes and 11 CCA-derived components. Some semantic metadata tags were particularly learnable (for example "Medical Use", "Mystical Experiences", "Alone", "Addiction Habituation") with ROC AUCs ranging from 0.88 to 0.95 for the better-performing tags. Model confusions tended to mirror biochemical and pharmacological relationships (e.g. phenethylamines vs tryptamines are more likely to be misclassified for one another than for benzodiazepines or opioids). BERTowid also produced gender classification with ROC AUC 0.85 and age regression with Pearson correlation 0.56 despite missingness and class imbalance in self-reported demographics. The ability to recover sensitive features was noted as a step towards debiasing. BERTiment generalized emotion detection to the testimonial domain: emotion ROC AUCs on unseen data ranged from 0.72 for "Realization" to 0.97 for "Love". A hedonic-tone classifier trained on IMDB reviews provided a complementary axis that sorted emotions by valence: e.g. "Admiration", "Pride", "Love" correlated positively; "Annoyance", "Disgust" correlated negatively. Clinical validation consisted of a psychiatrist manually adjudicating 393 emotions from 256 Erowid excerpts; human labels were within the model's top 10 predicted emotions 87% of the time, top 5 73% and top 1 42%, reported to be within inter-human variability from the original GoEmotions resource. Temporal sentiment trajectories revealed pharmacologically meaningful patterns via DTW. MDMA and the entactogen MDA showed high and increasing trajectories for "Love", distinguishing them from other drugs. Stimulants (cocaine, amphetamine, methamphetamine) exhibited gradually rising "Sadness" trajectories, while antidepressants (paroxetine, venlafaxine, sertraline) started higher on sadness but declined over the testimonial. Tryptamines (notably DMT and 5-MeO-DMT) showed elevated "Surprise", "Curiosity" and "Realization" relative to phenethylamines, which leaned more to "Admiration", "Excitement" and "Gratitude". Salvinorin A (a diterpenoid) shared high rankings on "Surprise", "Curiosity", "Realization" and also ranked highly on "Mystical Experience", paralleling short-acting tryptamines. Opioids were lower on curiosity-related emotions but higher on "Relief". Across drugs, "Neutral" sentiment typically decreased over the course of a testimonial while "Optimism" rose sharply near the end, resembling a peak-end effect. CCA uncovered 11 statistically significant components that linked word-usage to receptor-affinity profiles. The dominant component (CCA 0) contrasted a pole of abstract, transcendental and perceptual language (terms such as "reality", "universe", "peak", visuals, music) associated with DMT, LSD and psilocin and serotonergic affinities (5-HT2A, 5-HT1A, 5-HT2C) with expression concentrated in medial prefrontal cortex, against a pole of mundane suffering words ("depression", "pain", "addiction", "work") associated with stimulants and oxycodone and affinities at MOR, DOR, NET and DAT with expression in posterior cingulate and inferior parietal lobule. CCA-derived components often produced spatially contiguous cortical maps and bilateral symmetry when receptor weights were projected using Allen Brain Atlas gene-expression. BERTowid could predict CCA component weights from testimonial excerpts with Pearson correlations that declined with component order (for example 0.68 for CCA 0 down to 0.24 for CCA 11), indicating components explaining more variance were easier to learn from text. Overall, the three methods (BERTowid, BERTiment and CCA) converged on a broad dichotomy between "mystical/expansive" experiential language tied to serotonergic psychedelics and a "mundane/suffering" pole tied to stimulants and certain psychiatric medications. MDMA in particular occupied a uniquely positive affective niche across models, aligning with clinical interest in its therapeutic potential.

G. and colleagues interpret their findings as evidence that combining linear CCA and nonlinear transformer models can produce unified, biochemically anchored and temporally sensitive representations of psychoactive experiences from retrospective text. They emphasise that these language-derived representations capture clinically relevant qualities — the intensity of mystical experience, the depth of joy, anger and grief — and that temporally resolved trajectories reflect known pharmacological distinctions and anecdotal phenomenology reported by psychonauts and researchers. The authors situate their results alongside prior NLP analyses of Erowid, noting agreement on associations such as antidepressants being linked to negative affect (which they partly attribute to ascertainment bias). They also highlight that their transformer-based construction of detailed temporal trajectories is a novel contribution beyond earlier bag-of-words or topic modelling approaches. The CCA findings, in particular the dominant component contrasting lucid, transcendental phenomena with somatic suffering and addiction-related themes, are presented as reinforcing the transformer-derived emotion landscapes. Limitations acknowledged by the authors include inherent noise and biases in crowd-sourced retrospective reports (uncertainty about dosage, chronology, impurities and effects on memory), and the interpretive limits of receptor affinity data (affinity does not indicate functional effects such as agonism versus antagonism or biased signalling). They note ascertainment bias in the dataset may confound associations between drug classes and affective language. For future research and application, the authors propose combining these linguistic trajectories with concurrent high-temporal-resolution modalities (EEG, ECG, fMRI) to build cross-modal representations of acute drug states and to enable real-time neurofeedback. They suggest such combined representations could aid personalised, responsive psychoactive sessions in clinical practice and that transformer models’ zero-shot capabilities might be harnessed to monitor and guide emotional dynamics during therapy. The authors present these applications as plausible next steps while recognising that prospective clinical outcome data and cross-modal validation would be required to establish therapeutic utility.