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

Forum topic · 小凯 · 2026-08-13

Summary

ConVAWG is a retrieval-grounded framework for generating CPS-aligned synthetic multi-turn dialogues that model Violence Against Women and Girls (VAWG) scenarios. Presented in arXiv paper 2608.11200 by Chen Lyu and colleagues, the work addresses the scarcity of real conversational data in sensitive domains, where privacy and legal constraints prevent large-scale dataset release and prior research has largely focused on sentence-level online toxicity. ConVAWG constructs scenarios from character seeds, ONS-reported demographic patterns, official crime definitions, and retrieved domestic abuse review cases; converts them into hierarchical event timelines; generates multi-scenario role-played dialogues; and applies targeted activation-steering toxicity control to appropriate utterances. The authors release over 6,000 multi-turn conversational events across 200 scenarios with rich scenario-level, event-level, and turn-level metadata. Extensive human evaluation, LLM-as-Judge assessments, ablation studies, and downstream tasks demonstrate strong dialogue quality and domain fidelity.

Paper Overview

Field: NLP Authors: Chen Lyu, Xingwei Tan, Simon Cullen, Shelley Wilson, Lois Arthurs, Arshad Jhumka, Gabriele Pergola Published: 2026-08-11 arXiv: 2608.11200

Abstract

Synthetic dialogue generation offers a way to study conversational dynamics in sensitive domains where real data are difficult to access, release, or annotate. The underlying abuse may occur online or offline: threats and coercion can appear directly in messages, while behaviours such as surveillance, isolation, stalking, and physical violence may be planned, disclosed, or referred to conversationally. Privacy and legal constraints make it difficult the release of large-scale real conversation datasets; existing work has mostly focused on sentence-level toxicity of online abuses, leaving a gap in modelling abuse as a relational and temporally unfolding phenomenon. In this work, the authors focus on modelling Violence Against Women and Girls (VAWG) scenarios as multi-turn dialogues.

Key Contributions

ConVAWG framework: a retrieval-grounded pipeline for generating CPS-aligned synthetic VAWG chat dialogues.
Grounded scenario construction: scenarios are built from character seeds, UK Office for National Statistics (ONS) demographic patterns, official crime definitions, and retrieved domestic abuse review cases.
Structured generation: scenarios are converted into hierarchical event timelines, then rendered as multi-scenario role-played dialogues.
Controlled toxicity: targeted activation-steering toxicity control is applied to appropriate utterances.
Released dataset: over 6,000 multi-turn conversational events across 200 scenarios, with rich scenario-level, event-level, and turn-level metadata.

Evaluation

Extensive human evaluation, LLM-as-Judge assessment, ablation studies, and downstream task benchmarks demonstrate strong conversational quality and domain fidelity.

--- *Auto-collected on 2026-08-13*