Detecting Novel and Emerging Drug Terms Using Natural Language Processing: A Social Media Corpus Study

Area of Science:

• Computational Linguistics
• Data Science
• Epidemiology

Background:

• The emergence of new psychoactive substances (NPS) and evolving traditional drug use patterns challenge researchers and public health officials.
• Accurate tracking of drug terminology requires understanding terms used by drug users, necessitating innovative data sources.
• Social media analysis offers a promising avenue for near real-time discovery and monitoring of drug-related language.

Purpose of the Study:

• To evaluate the feasibility of employing distributed word-vector embeddings trained on social media data.
• To identify previously unrecognized drug terms by analyzing semantic relationships in online discourse.
• To assess the potential for automated drug term discovery systems.

Main Methods:

• A continuous bag of words (CBOW) model was trained on a large Twitter dataset (approx. 884.2 million tokens) from July 2016.
• Word embeddings were queried for terms semantically similar (cosine similarity) to known marijuana slang.
• Generated candidate lists were compared against expert-curated lists to validate novel terminology.

Main Results:

• The method generated 200 candidate terms for marijuana, with 115 confirmed as relevant (65 substance terms, 50 paraphernalia terms).
• Thirty novel terms, not present in expert lists, were successfully identified, indicating successful discovery of new drug language.
• Some novel terms appeared to have emerged only 1-2 months prior to the data collection period.

Conclusions:

• The approach shows significant promise for discovering novel drug terms, evidenced by the identification of 30 new marijuana-related terms from one month of data.
• While human review remains necessary due to current precision limitations, this pilot study serves as a proof of concept.
• This work represents a foundational step towards developing automated systems for real-time tracking of emerging NPS terminology.