Quantitative natural language processing markers of psychoactive drug effects: A pre-registered systematic review

A systematic review of studies using natural language processing to analyze speech and text after psychoactive drug use found that all studied substances—stimulants, MDMA, cannabis, ketamine, and psychedelics—alter language production. Emerging patterns include increased verbosity with stimulants, reduced lexicon with LSD, increased semantic proximity to emotional words with MDMA, increased positive sentiment with psilocybin, and altered speech acoustics with cannabis. Only one study provided externally validated support for identifying MDMA intoxication using NLP and machine learning. Meta-analysis was not possible due to heterogeneity and few studies. The authors call for standardized speech tasks and a shared language corpus to improve replicability.

Study at a glance

Abstract

Psychoactive substances used for recreational purposes have mind-altering effects, but systematic evaluation of these effects is largely limited to self-reports. Automated analysis of expressed language (speech and written text) using natural language processing (NLP) tools can provide objective readouts of mental states. In this pre-registered systematic review, we investigate findings from applying the emerging field of computational linguistics to substance use with specific focus on identifying short-term effects of psychoactive drugs. From the literature identified to date, we note that all the studied drugs – stimulants, 3,4-methylenedioxymethamphetamine (MDMA), cannabis, ketamine and psychedelics – affect language production. Based on two or more studies per substance, we note some emerging patterns: stimulants increase verbosity; lysergic acid diethylamide reduces the lexicon; MDMA increases semantic proximity to emotional words; psilocybin increases positive sentiment and cannabis affects speech stream acoustics. Ketamine and other drugs are understudied regarding NLP features (one or no studies). One study provided externally validated support for NLP and machine learning-based identification of MDMA intoxication. We could not undertake a meta-analysis due to the high degree of heterogeneity among outcome measures and the lack of sufficient number of studies. We identify a need for harmonised speech tasks to improve replicability and comparability, standardisation of methods for curating and analysing speech and text data, theory-driven inquiries and the need for developing a shared ‘substance use language corpus’ for data mining. The growing field of computational linguistics can be utilized to advance human behavioral pharmacology of psychoactive substances. Achieving this will require concerted efforts towards consistency in research methods.