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

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

Summary

This paper introduces ConVAWG, a retrieval-grounded framework for generating controlled synthetic chat dialogues in the domain of Violence Against Women and Girls (VAWG). The work addresses the challenge of studying conversational dynamics in sensitive domains where real data are difficult to access, release, or annotate due to privacy and legal constraints. The authors note that existing work has largely focused on sentence-level toxicity of online abuses, leaving a gap in modelling abuse as a relational and temporally unfolding phenomenon. ConVAWG models VAWG scenarios as multi-turn dialogues, capturing abuse as it unfolds across turns, speakers, relationships, events, and gradual temporal escalation.

The framework consists of four stages. Stage 1 constructs scenario specifications and narrative outlines from PersonaHub victim persona seeds, sampled to match UK Office for National Statistics (ONS) victim statistics, and grounded in Crown Prosecution Service (CPS) crime definitions and retrieved Domestic Homicide Review (DHR) cases. Stage 2 converts each scenario into a hierarchical event graph of timestamped sub-events, spacing early interactions sparsely and later high-risk ones more densely. Stage 3 realizes these events as online chat scripts and role-plays them into multi-turn dialogues with cross-scene continuity and retrieved style notes. Stage 4 applies activation-steered rewriting only to selected perpetrator utterances associated with escalation, adjusting toxicity intensity without disrupting the surrounding conversation.

The resulting dataset contains over 6,000 multi-turn dialogue events across 200 scenarios, with rich scenario-, event-, and turn-level metadata. The dataset covers 18 CPS-aligned VAWG crime types, with an average of 8.7 labels per scenario, reflecting the frequent co-occurrence of abuse forms in real-life VAWG cases. Each scenario includes persona profiles, relationship information, crime labels, hierarchical events, sub-events, and corresponding multi-turn dialogues, with an average of 31 dialogues per scenario and 15 turns per dialogue.

The authors conduct extensive experimental evaluation through human assessment, LLM-as-Judge evaluation, ablations, and downstream tasks. Human evaluation uses six female annotators with high inter-annotator agreement (mean Gwet's AC2 of 0.926). Four LLM judges (GPT-5.2, Gemini-3.1, DeepSeek-V4, and Qwen3.5) achieve strong alignment with human consensus (mean AC2 ≥ 0.925). ConVAWG achieves the best overall score (Avg 4.75) compared to eight baselines, including five direct-generation baselines (LLaMA-3.1-70B-Instruct, Qwen3.5-35B-A3B, Gemini-3.1-Flash-Lite, DeepSeek-V4-Flash, and GPT-5.2) and three framework baselines (DiaSynth, SynDG, and SPASM). ConVAWG leads most clearly on Humanlikeness (4.51 vs. 4.34) and Toxicity Realism (4.93). Statistical significance tests (Holm-corrected paired Wilcoxon tests) place ConVAWG significantly above every baseline on both Avg and Avg-D (all p<10−4).

Ablation studies show that removing style notes reduces overall average scores by 0.16 and 0.13 points under two judges, with the largest degradation in Persona Consistency (−0.65). Disabling CAA steering sharply reduces Detoxify toxicity scores from 0.144 to 0.038 while LLM-judge averages remain largely stable, indicating that CAA mainly controls explicit toxic wording without substantially degrading overall dialogue quality. Backbone model ablation shows GPT-5.2 achieves the highest overall score and was chosen as the backbone model.

Downstream utility experiments distinguish controllability validation from external utility. For controllability, fine-tuned encoders recover toxic behaviour classification (0.782 macro-F1) and victim-perpetrator relationship prediction (0.629 weighted F1) effectively. For external utility, escalation forecasting models detect escalation jumps substantially better than baselines (0.548 jump F1), and cross-corpus transfer to Conversations Gone Awry corpora shows above-chance performance (best AUROC 0.643). Real-data transfer experiments with MentalManip show that a detector trained only on MentalManip identifies ConVAWG dialogues as manipulative (0.755 macro-F1), and pre-training on ConVAWG before fine-tuning on 500 real examples matches real-only performance on MentalManip while substantially outperforming it on ConVAWG.

The authors acknowledge limitations, including the UK-specific nature of grounding sources (CPS definitions, ONS statistics, DHR reports), which may transfer imperfectly to other jurisdictions. The framework captures behaviours only when they produce a conversational trace, with offline or behavioural abuse primarily encoded in hierarchical event graphs and metadata. Ethical considerations note that no private conversations or identifiable personal communications are included, all dialogues are fictional, and access to the archival release is governed by a data use agreement restricting use to non-commercial research and prohibiting harassment, profiling of real individuals, training systems primarily intended to generate abusive content, and application to real-world cases.

Improvements to AI Systems:

1. Temporally-Aware Abuse Detection Models

• Train models to recognize abuse as a multi-turn escalation pattern rather than isolated toxic sentences.

• Use the hierarchical event graphs to predict escalation risk (e.g., from early sparse interactions to dense high-risk events).

• Improved system: A chatbot moderator that flags conversations before explicit abuse occurs by detecting relational tension, power imbalances, and gradual intensity shifts across turns.

2. Controllable Toxicity Generation with Activation Steering

• Implement activation-steered rewriting (CAA) to adjust toxicity intensity selectively on perpetrator utterances without degrading dialogue coherence.

• Improved system: A safe role-play simulator for training law enforcement or social workers, where they can practice de-escalation by adjusting the aggressiveness of a virtual perpetrator in real-time while keeping the conversation natural.

3. Cross-Domain Transfer for Sensitive-Domain Dialogue Systems

• Use ConVAWG’s retrieval-grounded framework (scenario specs → event graphs → dialogue realization) to generate synthetic data for other privacy-restricted domains (e.g., elder abuse, domestic violence in non-UK jurisdictions, cyberbullying).

• Improved system: A data augmentation pipeline that produces labeled, multi-turn dialogues for domains where real data cannot be shared, enabling fine-tuning of detectors without privacy violations.

4. Style-Grounded Persona Consistency

• Incorporate retrieval of style notes (e.g., victim-perpetrator communication patterns) to maintain persona consistency across long dialogues.

• Improved system: A virtual interviewer for mental health or legal intake that adapts its language and questioning style to match the victim’s communication profile, improving rapport and information elicitation.

5. Fine-Grained Multi-Label Crime Co-Occurrence Prediction

• Leverage the 18 CPS-aligned crime types and average 8.7 labels per scenario to train multi-label classifiers that capture co-occurring abuse forms (e.g., coercive control + physical assault).

• Improved system: An automated case triage tool for helplines that identifies multiple overlapping abuse types from a user’s narrative, enabling more accurate referral to specialized services.

6. Escalation Forecasting for Early Intervention

• Train models on the timestamped event graphs to predict when a conversation is likely to jump to a higher-risk event (e.g., from verbal abuse to threats of violence).

• Improved system: A real-time alert system for online platforms that warns moderators or provides resources to users when a conversation trajectory matches known escalation patterns, before physical harm occurs.

7. Cross-Corpus Robustness for Toxicity Detection

• Use ConVAWG’s transfer experiments (e.g., to Conversations Gone Awry) to build detectors that generalize across different conversational contexts and platforms.

• Improved system: A unified moderation API that works across social media, dating apps, and support forums, with consistent performance despite varying dialogue styles and user demographics.

8. Ethically-Constrained Synthetic Data Generation

• Adopt the framework’s data use agreement and fictional-dialogue design to generate synthetic datasets that avoid privacy violations while retaining realistic relational dynamics.

• Improved system: A research tool that produces shareable synthetic corpora for academic study of sensitive topics (e.g., human trafficking, stalking) without exposing real victim data, accelerating scientific progress while protecting vulnerable populations.