ConVAWG: A Retrieval-Grounded Framework for Controlled Synthetic Dialogue Generation in Violence Against Women and Girls

Imagine you're trying to understand the dynamics of conversations around sensitive topics like violence against women and girls. Gathering real data in these areas is incredibly challenging due to privacy concerns and the potential for harm, which means researchers often rely on limited datasets that don't capture the complexity of these interactions. Current methods typically focus on isolated instances of abusive language rather than the ongoing relational dynamics that unfold over multiple exchanges, which is a significant gap in understanding how such abuse manifests in conversation. This is what's called a failure to model the temporal and relational aspects of dialogue. To address this, the authors developed a framework called ConVAWG, which generates synthetic dialogues that reflect real-world scenarios of violence against women and girls. By using persona seeds and demographic data, they create rich, multi-turn dialogues that simulate the complexities of these interactions. The framework not only generates dialogues but also incorporates mechanisms to control for toxicity in the generated content, ensuring that the dialogues are both realistic and ethically sound. Compared to previous work, ConVAWG offers a more nuanced approach to studying conversational abuse by providing a large dataset of over 6,000 dialogue events across various scenarios. This means that for those building tools or conducting research in this area, there is now a more robust resource to explore the dynamics of abusive conversations, ultimately leading to better understanding and potential interventions.

The approach introduces a novel framework for generating synthetic dialogues in a sensitive domain.

The evaluation includes extensive human assessments and various metrics, indicating solid reliability.