A Survival Guide for Iranian Women Prescribed by Iranian Women: Participatory AI to Investigate Intimate Partner Physical Violence in Iran

Adel Khorramrouz, Mahbeigom Fayyazi, Ashiqur R. KhudaBukhshIJCAI 2024

Introduction

Intimate Partner Violence (IPV) remains a critical global issue, particularly in regions like Iran where documented gender inequality and social stigma create significant barriers for victims seeking help. Traditional data collection is hampered by legal and cultural constraints, leading researchers to turn to digital "safe spaces." This study focuses on Ninisite, a prominent Iranian parenting forum that serves as a vital hub for women to share experiences and seek advice. By employing a participatory AI framework, the researchers analyzed millions of comments to understand how the community responds to disclosures of physical abuse.

Methodology: Participatory AI and Rare Class Mining

Analyzing domestic violence on a massive forum like Ninisite presents a "needle in a haystack" problem. To address this, the researchers utilized a Participatory AI approach, involving an Iranian human rights advocate as a lead annotator to define culturally nuanced categories of social response.

The technical pipeline involved several sophisticated steps:

  1. Rare Class Mining: Using nearest neighbor sampling in document embedding spaces to identify threads likely discussing Intimate Partner Physical Violence (IPPV).
  2. LLM Filtering: Leveraging GPT-3.5 to filter irrelevant content, followed by human verification.
  3. Classification: Training ParsBERT, a transformer model specialized for Persian, to categorize community responses.
  4. Active Learning: Iteratively improving model accuracy by labeling the most informative examples to overcome extreme class imbalance.

The "Confront vs. Conform" Framework

A central contribution of this work is the categorization of community responses into three distinct buckets:

  • Empathy: Passive support without specific advice.
  • Conform: Urging the victim to endure, adjust, or "fix" the situation through submission or imagination.
  • Confront: Encouraging the victim to seek legal recourse, retaliate, or leave the relationship.
An example of a post describing physical abuse and a corresponding 'Conform' response advising the victim to imagine her husband is the 'best man in the world'.

Key Results and Societal Shifts

The study revealed that IPPV posts generate significantly higher engagement than standard forum topics, averaging 46.34 comments compared to 26.29. More importantly, the data suggests a measurable shift in societal attitudes.

The Rise of Agency

Contrary to stereotypes of passive endurance, "Confront" responses were more prevalent than "Conform" advice. Between 2018 and 2022, the proportion of "Confront" responses rose from 28.2% to 35.2%, indicating a growing trend of women encouraging each other to seek agency and challenge abuse.

Victim and Perpetrator Dynamics

Human annotation confirmed that while husbands are the primary perpetrators, the patriarchal structure often involves in-laws and fathers.

Distribution of perpetrators and victims identified in the Ninisite dataset, showing husbands as the primary perpetrators and wives as the primary victims.

The analysis also noted a small subset of "male victims," often occurring in contexts where women used physical force in self-defense or counter-attack.

Conclusion and Future Directions

While the study provides a powerful map of domestic violence discourse, it is limited by potential censorship on Ninisite and a focus on binary gender roles due to legal restrictions in Iran. Future research aims to broaden this "survival guide" by investigating non-physical forms of abuse, such as emotional and financial control, to provide a more comprehensive understanding of the challenges faced by Iranian women.

Summary updated

BibTeX
@inproceedings{khorramrouz2024survival,
  author = {Adel Khorramrouz and Mahbeigom Fayyazi and Ashiqur R. KhudaBukhsh},
  title = {{A Survival Guide for Iranian Women Prescribed by Iranian Women: Participatory AI to Investigate Intimate Partner Physical Violence in Iran}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure