Introduction
Online grooming is a predatory process characterized by gradual manipulation through multi-stage communication. Effective protective systems must intervene early, ideally before grooming escalates to physical harm. However, current Early Sexual Predator Detection (eSPD) models face a significant hurdle: "weak supervision." Most available datasets provide only a single label for an entire conversation (e.g., "Risky" or "Normal").
Because risky utterances are often sparse within long dialogues, models trained on these coarse labels tend to overfit. They often flag innocent messages as predatory simply because they appear in a conversation that eventually becomes harmful. This leads to high false-positive rates in normal interactions, rendering such systems impractical for real-world deployment.
Methodology: The SCoRL Framework
To address these limitations, we introduce SCoRL (Speed Control Reinforcement Learning), a framework designed to capture turn-level dynamics. SCoRL moves beyond binary chat labels by optimizing for both detection speed and accuracy at the individual message level.
1. Turn-Level Labeling
Utilizing Luring Communication Theory (LCT), individual messages are categorized into specific stages. Turns are only marked as "Risky" if they fall into the "Grooming" (introduction of sexual topics) or "Approach" (solicitation of contact) phases.
2. Two-Stage Training Process
The SCoRL framework operates in two distinct phases:
- Supervised Fine-Tuning (SFT): A RoBERTa-based model is initialized to predict turn-level risk labels based on the LCT annotations.
- Reinforcement Learning (RL): The model is treated as an agent in a Markov Decision Process (MDP). At each turn, it must decide whether to "continue" the dialogue or "stop" and trigger an alert.
3. Reward Function and the "Wait and See" Strategy
The core innovation of SCoRL is its reward function, which balances the trade-off between speed and accuracy. The agent receives a "speed" reward for correctly stopping on a risky turn (higher for earlier detection) and is penalized for premature stops on non-risky turns. Crucially, the model is also rewarded for "waiting" until the end of a normal conversation without triggering a false alarm. This "wait and see" strategy allows the model to handle "hard negatives"—such as consensual adult cybersex—where sexual language may be present but the context is not predatory.
Experimental Setup
The study utilizes the PANC dataset, a combination of the PAN12 shared task and ChatCoder2 (CC2). CC2 provides the essential turn-level strategy annotations based on LCT. The dataset includes both grooming chats and normal segments, including difficult-to-distinguish "hard negatives."
Results and Discussion
SCoRL demonstrates a substantial improvement over existing state-of-the-art baselines, such as SBERT-base. Key performance metrics include:
- Latency-F1 Score: SCoRL achieved 0.365, a significant leap from the baseline's 0.062.
- Precision: Precision increased dramatically to 0.475 (compared to 0.047), indicating a much lower rate of false alarms.
- Success Rate: In identifying specific risky utterances (Grooming or Approach), SCoRL reached an 85.5% success rate, while the baseline managed only 20.29%.
- Efficiency: With an inference time of just 23ms per turn on standard hardware (e.g., RTX 3090), the model is well-suited for real-time monitoring.
By anchoring detection in established communication theory (LCT), SCoRL identifies specific predatory strategies rather than merely flagging "dirty words." This grounding ensures that the model's detections are contextually relevant and theoretically sound.
Conclusion and Future Work
SCoRL represents a significant advancement in the early detection of online predators by shifting the focus from chat-level to turn-level optimization. While the results are promising, future work will focus on expanding the limited training data (currently 77 annotated conversations) and incorporating the victim's perspective to provide a more holistic view of predatory interactions. Additionally, enhancing the model's ability to handle non-linear strategy shifts remains a key challenge for future research.
Summary updated