Analysing Valence Shifts in Scientific Narratives on Psychedelics using BERT and Topic Modeling.

Throughout the years, psychedelic substances such as psilocybin and ketamine have gained attention in neuroscience and psychiatry due to their potential therapeutic effects on treatment-resistant depression. However, academic literature discussing these substances remains heterogeneous in tone and often reflects both cautious optimism and critical concern of abuse and risk. This thesis presents a computational framework for analysing how the valence and sentiment in scientific narratives shift depending on whether psychedelics, such as psilocybin and ketamine, are discussed in therapeutic context or as substances of abuse. A fine-tuned version of the SciBERT language model was employed to predict sentiment at different granularities across a corpus of academic articles. To capture narrative shifts, sentiment scores were aggregated by article sections and publication year, revealing trends of increasing positivity in recent publications on psilocybin, and a more stable sentiment trajectory in ketamine articles. BERTopic was also applied for topic modeling, enabling unsupervised and semi-supervised search of dominant themes. Seeded topic modeling allowed identification of paragraphs centred on pre-defined themes, providing a path into the framing of psychedelics as either beneficial treatments or abusive substances. The relationship between sentiment scores and topic dominance was examined, revealing complex interactions not always aligned with intuitive expectations. Together, sentiment analysis and topic modeling offer a novel methodology for mapping valence trends and thematic structures in scientific discourse. The results suggest that sentiment not solely dictated by topic but reflects nuances and cautious formulations across sections and time. This approach contributes to the developing field of computational narrative analysis and provides a tool for tracing evolving framings of controversial biomedical topics in academic literature.