Lived experiences of substance use and expressions of stigma
TL;DR
This dissertation advances computational social science and NLP literature by providing robust computational approaches for understanding the lived experiences of people who use drugs, specifically how stigma manifests in their online disclosures and demonstrating how computational methods can effectively model stigma as a dynamic process.
Abstract
The stigma surrounding substance use and substance use disorders (SUD) represents a significant barrier to behavior change (e.g., help-seeking, recovery), with only 7% of individuals with SUD receiving any form of help. This dissertation focuses on understanding how stigma manifests in online communities and developing computational interventions to address it. This research has three primary aims to: (1) characterize online substance use communities and expressions of stigma, (2) model internalized stigma as a psychosocial and temporal process, and (3) develop and deploy narrative-aware natural language processing (NLP) tools for the detection and transformation of stigmatizing language. To achieve these aims, we conducted network analyses of drug-related subreddits to identify influential communities and developed a multi-level taxonomy for classifying personal drug experiences disclosed on Reddit. Building on these foundations, we identified distinct phenotypes of stigma expressions using computational methods and validated them against established theoretical frameworks. The next phase of the research involves modeling the internalization of stigma over time. Together, these descriptive and temporal characterizations motivate the final component: creating and evaluating computational tools to intervene in stigmatizing communication. These tools include a framework to transform stigmatizing language into more empathetic alternatives, and a persona-aligned empathetic response system developed across three phases: (i) deriving narrative archetypes of self-stigma from publicly available Reddit data, (ii) building a classification system to recover these personas in order to simulate a cold-start chat environment, and (iii) qualitatively analyzing the responses of a narratively-aware chat system against baseline systems. This dissertation advances computational social science and NLP literature by (1) providing robust computational approaches for understanding the lived experiences of people who use drugs, specifically how stigma manifests in their online disclosures, (2) demonstrating how computational methods can effectively model stigma as a dynamic process, and (3) delivering practical, empirically-informed tools for addressing harmful language and supporting individuals affected by substance use stigma.