Giving Voice to Survivors: Cross-National Insights into Gender-Based Violence Recovery Using Natural Language Processing

• In Event: Advancing Racial Equity in Health and Social Policy: Evidence from Programs, Providers, and Policy Design

Abstract

Introduction
Gender-based violence (GBV) is a global crisis, affecting one in three women in their lifetime. As digital platforms become increasingly central to trauma recovery—especially for those without access to in-person services—there remains a gap in understanding what survivors want from these tools. Most evaluations are practitioner-led and narrowly outcome-focused, often overlooking survivor-defined healing goals.

This study analyzes over 10,000 open-text responses from survivors who used Bloom, a free online recovery program by Chayn, a global nonprofit led by survivors. The survey’s core prompt—“What’s one thing that you’re hoping Bloom will help you to understand or do?”—was intentionally open-ended, allowing survivors to define healing in their own words. This aligns with trauma-informed and feminist AI principles, but introduces analytic complexity due to wide variation in tone, structure, and cultural reference.

Purpose
This study asks: What do survivors of GBV want from digital trauma recovery, and how do their narratives vary across emotional, cultural, and demographic contexts?

Aims include:
- Inform trauma-informed, survivor-centered policy and service design
- Advance ethical, feminist approaches to natural language processing (NLP)
- Develop an open-source NLP pipeline to support future research and GBV programs

Methods
The dataset—one of the largest of its kind—spans multiple countries, identities, and neurocognitive experiences. A mixed-methods NLP pipeline was developed with Chayn and the Equity and Access in Algorithms, Mechanisms, and Optimization (EAAMO) working group. Components include:

- Rule-based classification of healing stages using a simplified Transtheoretical Model
- Thematic coding for trauma-informed perspectives (psychological, cognitive, emotional, behavioral)
- N-gram and clustering analysis to uncover survivor-defined themes and patterns
- Emotion detection using Hugging Face models trained on the EmpatheticDialogues dataset
- Ontology mapping to GBV-Heal domains and coping styles (problem-, emotion-, support-, meaning-focused)

All data was anonymized and handled under a signed NDA. Survivors were not consulted for supervised model training, so only unsupervised and rule-based methods were used. Non-English responses were translated with Google Translate and verified by multilingual Chayn experts.

Findings
Preliminary results show diverse healing goals and emotional tones. Some survivors were unsure how to begin, others expressed tentative hope, and many shared specific needs—often revealing deeper emotions not explicitly named. Tones ranged from flat or devastated to resilient and empowered. A common theme was the desire for peer connection, both to receive and give support.

The pipeline helped detect not just what survivors need, but how readiness and tone shift across healing stages—offering a scalable way to model emotional and cognitive complexity in recovery.

Conclusion
This project offers a survivor-informed, ethically grounded framework for analyzing GBV recovery narratives at scale. It demonstrates how NLP can preserve voice while producing insights that inform more responsive design.

Policy implications include:
- Culturally resonant messaging frameworks
- Peer-based, community-driven service models
- Timing-sensitive recovery interventions
- Open-source tools for ethical GBV evaluation across sectors

Author

• Ashley Khor, University of Pittsburgh
  Presenting Author