Real-time Chatbot Romance Scam Detection in Stage Classification

Motivation

• Romance scammers use fake online identities to gain victims’ affection and trust, then manipulate or defraud them through the appearance of intimacy.The introduction also notes victims may avoid disclosure or support because of embarrassment or fear of disapproval, with potential mental-health harms including depression, grief, or PTSD.Source: §I · p. 1

Study aim

• The study develops a chatbot for real-time romance-scam detection.The introduction frames this as a response to romance scams occurring on social media and the risks described above.Source: §I · p. 1

Stage models

• Andoh-Baidoo’s model divides romance scams into recruitment, grooming, initial extortion, and escalated extortion.The stages describe progression from recruiting people online through increasingly large financial demands or threats.Source: §II · p. 2
• Whitty’s seven-stage model covers ideal-partner seeking, fake profiles, grooming, extortion, repeated scams, sexual abuse, and re-victimisation.The stages span initial contact through further exploitation by purported investigators or refund agents.Source: §II · p. 2

Prevention approaches

• Prior prevention approaches include contextual analysis on dating platforms, NLP-based profile filtering, and combining images, profile descriptions, and structured data.Source: §II · p. 2

Study approach

• The study adopts Whitty’s model for its greater stage detail and uses a large language model to evaluate romance-scam stages.Source: §II · p. 2

A. Line Chatbot Real-time Detection

• A Line webhook receives messages, extracts user and authentication data, and sends text history for GPT-4o context analysis before stage classification.The system uses Python Flask and the Line Messaging API and informs users that messages are recorded and analyzed for privacy purposes.Source: A. Line Chatbot Real-time Detection · p. 2

B. Large Language Model in Stage Classification

• GPT-4o is prompted to identify experience versus dialogue, assign one of seven stages, extract scam-related triggers, and justify its choices with message evidence.Source: B. Large Language Model in Stage… · p. 3
• The classifier returns a stage, input type, labels, and rationale, with trigger keywords displayed to explain the classification.Examples of trigger categories include emotional manipulation, emotional triggers, economic extraction, trust, and geolocation.Source: B. Large Language Model in Stage… · pp. 3–4

C. Explanation Module

• GPT-4o generates explanations of the detected stage using stage details, triggers, input type, and the classifier’s rationale.The prompts request a brief explanation and an expanded explanation that includes how signals combine and any caveats or uncertainties.Source: B. Large Language Model in Stage… · p. 4

C. Prevention Advice Module

• The prevention module prompts GPT-4o for three practical suggestions and can expand each into implementation guidance; responses take a few seconds.The paper says the explanation module helps users understand the detected stage and the prevention module helps reduce scam risk.Source: B. Large Language Model in Stage… · p. 4

D. Visual Feedback in Interface Design

• The interface uses different colors for the seven stages and shows stage names and keyword icons for long-text and single-sentence inputs.The paper gives orange-yellow for stage 2 and red for stage 3 as examples; when no scam is detected, the chatbot returns a “no risk” flag.Source: A. Line Chatbot Real-time Detection · p. 3, D. Visual Feedback in Interface Design · p. 4

Stage evaluation

• 100% accuracy for “The sexual abuse” and 92.9% for “Continuation of the scam” were the highest reported stage accuracies.The paper attributes these results to clear cues such as sexual inducements or nude-photo requests and urgent assistance or money requests.Source: §IV · p. 4

The sting

• 28.6% accuracy was reported for “The sting”; all 20 misclassified cases were labeled “Continuation of the scam” in the no-context evaluation.Correct sting classifications included direct references to money or danger, while cases without direct information were confused with continuation.Source: §IV · p. 5

Discussion

• The authors state that real-world detection relies on context information and that the chatbot performs better in that setting than these context-free evaluation results suggest.Source: §IV · p. 5

• The chatbot provides real-time romance-scam detection in Line chatting as an application for reducing risk.Source: §V · p. 5
• Explanations of detected scam stages and prevention advice help users understand romance scams.Source: §V · p. 5
• The chatbot's explanations and advice can alert users to suspicious social-media chats.Source: §V · p. 5