Real-time Chatbot Romance Scam Detection in Stage Classification

arXiv:2610.20631 · cs.CL, cs.AI, cs.MM · Submitted 2026-09-30 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Real-time Chatbot Romance Scam Detection in Stage Classification".

Jane: The gist The study developed a chatbot to detect romance scams in real-time using seven stages from social science theories and large language models,

Tom: First, who's behind it and why it matters.

Paper summary: Tom: So, we're talking about this paper, "Real-time Chatbot Romance Scam Detection in Stage Classification." It’s about using AI to spot romance scams as they happen on social media. The authors are Yi-cheng Liu, Guan-ru Lai, and others from institutions like National Chengchi University and National Taiwan University.

Jane: That's right, Tom. The main idea here is that they built a chatbot that can detect romance scams in real time using the Line Messaging API as the way to get messages. They are using a large language model to figure out which stage of a scam it is based on seven stages from social science theory.

Lu: And what’s really interesting about their approach is that they use this large language model not just for detection, but also for explaining what kind of stage the scam is in and giving advice on how to stop it. They are using GPT-4o for both parts of that process <ref:2610.20631#pg1>.

Meng: I'm curious about how they handle the actual deployment. How fast can this thing actually work when you have a flood of messages? We need to know if this is something that runs smoothly at scale or if it just breaks when things get busy.

Lalam: Based on what I've reviewed, the system architecture involves context analysis and stage classification implemented in Python Flask and the Line Messaging API. They take user input texts, extract the user ID, texts, and token for message signature authentication.

Tom: Right, so they’re analyzing messages before processing them further for privacy concerns by recording text messages in user history. And they use GPT-4o to analyze that history for context analysis because the advantage of the large language model is that it can identify the semantic features of romance scams efficiently <ref:2610.20631#pg1>.

Jane: So, when it comes back to detection, their process involves the webhook receiver calling a message process to determine the stage of a romance scam. And they give users responses using a predefined template and a text style that shows them the stage number and name with triggered factors based on those seven stages.

Meng: How detailed are these explanations? Because if I’m shipping this, I need to know if the explanation is too wordy or if it gives us actionable intelligence fast enough for someone to use it while they're in a chat.

Lu: They do make the generated texts in explanation mode in paragraphs and the listing of prevention advice differs for each stage considering readability. Otherwise, when no romance scam is detected, they return a flag that says “no risk.”

Tom: And let's talk about those seven stages they use. They adopt the Whitty stage model for defining the stages of romance scam thirteen <ref:2610.20631#pg3>. The record of seven stages includes stage zero: motivated to find the ideal partner, stage one: presented with the ideal profile, stage two: presented with the ideal profile, and so on up to stage six which is re-victimisation <ref:2610.20631#pg3,stage 0: motivated to find the ideal partner, stage 1: presented with>.

Paper summary: Jane: That's a lot of specific labels. For example, Stage three is called "the grooming process," and Stage five is "continuation of the scam <ref:2610.20631#pg2>." They use GPT-4o for analyzing the scam stage of the Line messages using a very specific prompt input that asks it to decide between a personal experience description or pure dialogue first <ref:2610.20631#pg1>.

Meng: That prompt setup sounds pretty complex. What’s the actual cost or complexity when you run this whole thing in production? We're not just talking about training it once; we're talking about inference costs for every single message processed in real time.

Lalam: The prompt input for context analysis of user text history is “You are an expert romance-scam detection assistant based on Witty et al.’s seven-stage framework. Given a user message (which may be either a pure dialogue snippet or a user describing their personal experience), do the following: one <ref:2610.20631#pg1>. Decide whether the message is a personal experience description or pure dialogue.”

Tom: That prompt forces the model to first classify if it’s an experience versus pure dialogue before moving on to stage identification, which is a smart way to handle ambiguity. Then, it asks it to identify which scam stage fits best using a specific integer mapping from zero to six.

Jane: And once that stage is identified, the model also extracts all relevant scam-related labels or triggers present, like "payment," "urgency," or "money request." This gives them concrete things they can point back to when explaining the result.

Meng: So it's not just saying “this is stage five”; it’s saying “this is stage five because of the trigger factor money request and urgency.” That kind of detail would be useful for us to debug any failures in deployment.

Lu: Exactly, that's where the model provides the rationale, citing exact phrases or patterns from the message that justify its classification. This level of justification makes it much more trustworthy than just a simple stage label.

Tom: And when they move to generating explanations and prevention advice, they use GPT-4o again for both <ref:2610.20631#pg1>. For example, if it detects stage five the prompt asks it to provide two to three sentences briefly explaining why that judgment was made <ref:2610.20631#pg2>.

Jane: Then there’s a second prompt where they ask for more information. If you already have the stage and the triggers, you can ask for a detailed explanation of why that classification makes sense, including any caveats or uncertainties.

Meng: And then there’s the advice part. The final prompt asks it to provide exactly three practical prevention suggestions for that specific fraud stage based on those trigger factors, returned as a numbered list where each suggestion is in one sentence. That’s what we need to see for real-world utility.

Paper summary: Lalam: To give you the idea of the advice generation prompt, it’s “Given fraud stage five (Continuation of the Scam) and trigger factors trigger names, provide exactly three practical prevention suggestions <ref:2610.20631#pg2>. Return them as a numbered list, each in one sentence.”

Tom: The results they showed are pretty interesting too. They collected romance scam data from Reddit and labeled it manually with those seven stages, and their chatbot showed high accuracies for "The sexual abuse" at one hundred percent and "Continuation of the scam" at ninety-two point nine percent.

Jane: That's a solid metric for detection, especially when you look at what features define those stages. For instance, stage five has clear features like urgent assistance or money requests, and they classified those correctly.

Lu: But they also noted that wrong classifications happen when the data has obvious features of other stages without context information. For example, a message about someone sending over thirty thousand dollars over seven years was misclassified as stage six, re-victimization.

Meng: So the paper admits that relying only on the text without deep context can lead to errors, which is a fair limitation when you're trying to deploy this in a messy social media environment. It’s not perfect detection just because it uses an LLM.

Tom: The conclusion they draw is that this chatbot provides a useful application for reducing the risk of romance scams and can be applied in real scenarios for real-time detection right there in Line chatting. The explanations and the prevention advice also give people useful information to understand those stages and to alert them when they see suspicious chatting on social media.

Jane: So, putting it simply, this paper shows how combining a specific social science framework with a powerful large language model can create an automated system that doesn't just flag a scam but explains exactly what's happening and tells you how to stay safe.

Lu: The implication for the field is using these established seven-stage theories—like the Whitty model—as a structured input for LLMs to perform nuanced, context-aware classification in dynamic communication streams.

Meng: For production, it means we have a framework that maps ambiguous user text into actionable stages with justifications, which should help us build better filters than just keyword matching.

Lalam: I'm confident in the model’s ability to follow these instructions and connect the detection to the prevention strategies based on its training.

Tom: So, if you’re driving or cooking and you get a suspicious message, this framework is designed to give you a quick breakdown of what kind of scam it might be before it escalates. That's what we saw with "Real-time Chatbot Romance Scam Detection in Stage Classification."

Conclusion: Tom: So we've seen how this chatbot uses seven social science stages to spot romance scams in real time using Line messages and GPT-4o for explaining what happened <ref:2610.20631#pg1>.

Jane: That's right, Tom. The title of the paper is "Real-time Chatbot Romance Scam Detection in Stage Classification." And the authors are Yi-cheng Liu, Guan-ru Lai, and others from places like National Chengchi University and National Taiwan University.

Lu: What this means for the world is that we're moving beyond just looking for single keywords or simple patterns in a message. We’re building a system that understands the entire story of the scam based on established psychological models.

Meng: So, instead of just saying "this has a money request," it figures out if it fits stage four, "the sting," which tells us exactly what kind of manipulation is happening right then and there.

Lalam: The core contribution here is taking those complex seven stages and feeding them into a large language model to classify the message in real time. And once classified, the AI explains the specific trigger factors that led to that stage.

Tom: It’s about practical application, you know? This isn't just theory on paper; it’s a system designed to reduce risk for users right inside their Line chat environment.

Jane: Exactly. The implication is that platforms and security teams could use this framework to create much more nuanced warning systems than they have now. It helps people understand the scam process itself, not just avoid a single suspicious link.

Lu: But the paper did flag a limitation, which is important for us to remember. The detection accuracy relies heavily on having enough context in that message history for the model to make a correct stage classification.

Meng: Right, so if someone sends a very short message with no history, that system might struggle to be sure about the stage. We need more data variety for it to be robust across all possible social media interactions.

Tom: It shows how combining established psychological frameworks with modern AI tools can give us a tool that actually helps people navigate these kinds of digital dangers safely.

Jane: And now we’re going to look at how they specifically trained that model and what data they used to make it so accurate in those real-time classifications.

National Chengchi University · Graduate Institute of Artificial Intelligence Cross-disciplinary Technology, National Taiwan University of Science and Technology · Institute of Information Science, Academia Sinica · Interdisciplinary Artificial Intelligence Center, National Chengchi University

cs.CL, cs.AI, cs.MM

Submitted: 2026-09-30

Updated: 2026-09-30

Comments: Accepted in Taiwan Academic Network Conference (TANET 2026)

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 69/100

The gist: The gist The study developed a chatbot to detect romance scams in real-time using seven stages from social science theories and large language models, providing explanations and prevention advice for

Key concepts

Seven Stages Model
This is a framework from social science used to categorize romance scams into seven distinct phases, ranging from initial interest in a partner to re-victimization. The chatbot uses this model to systematically analyze user messages and determine where the scam currently stands.
Line Messaging API
This is the technical tool used for the chatbot's real-time operation. It allows the system to receive, process, and send messages directly within a Line chat environment, enabling instant detection when a user interacts with it.
GPT-4o Model
A powerful large language model used for complex analysis. It is employed to understand the semantic features of romance scams within text and to generate nuanced explanations about why a specific message fits a certain scam stage, along with relevant prevention tips.

Summary

The gist The study developed a chatbot to detect romance scams in real-time using seven stages from social science theories and large language models, providing explanations and prevention advice for users.

System Architecture and Detection

The system architecture for the chatbot romance scam detection in real-time includes context analysis and stage classification implemented in Python Flask and Line Messaging API From the Line messages of user input, given an input of texts, the webhook receiver on local device receives the Line webhook and the texts, and extracts the user ID, texts and token for message signature authentication The system notices the users of recording and analysis of texts messages in advance for privacy concern Using the record of text messages in user history, the Generative Pre-trained Transformer 4 Omni model (GPT-4o) model was adopted for context analysis, and the advantage of the LMM model is that the semantic features of romance scam can be identified efficiently According to the results of context analysis, the webhook receiver call the message process to determine the stage of the romance scam in stage classification GPT-4o model was also adopted for providing explanations and prevention advice to the users Chatbot responses to the users using a predefined template and a text style in the form of a stage number, a stage name, and a list of triggered factors according to the seven stages with brief descriptions in the Line chatting In addition, the generated texts in explanation mode are in paragraphs, and the listing prevention advice differs in different stages both considering the readability Otherwise, when no romance scam is detected, a flag of “no risk” is returned

Stage Classification Methodology

The study adopts the Whitty stage model for defining the stages of romance scam [13] The record of 7 stages includes the stage ID (0 to 6) as follows: stage 0: motivated to find the ideal partner, stage 1: presented with the ideal profile, stage 2: presented with the ideal profile, stage 3: the grooming process, stage 4: the sting, stage 5: continuation of the scam, and stage 6: re-victimisation (re-victimization) GPT-4o model was adopted for analyzing the scam stage of the Line messages The prompt input for context analysis of user text history is as follows: “You are an expert romance-scam detection assistant based on Witty et al.’s seven-stage framework. Given a user message (which may be either a pure dialogue snippet or a user describing their personal experience), do the following: 1. Decide whether the message is a personal experience description or pure dialogue. (experience vs dialogue) 2. Identify which scam stage the message best fits into. Use the integer mapping: 0: Find the Dream Mate 1: Contact via fake profile 2: Grooming 3: The Sting 4: Continuation of the Scam 5: Sexual Exploitation 6: Revictimization. Use these labels to extract all relevant scam-related labels/triggers present (e.g., payment, urgency, identity inconsistency, unverified communication, romance, authority, crisis, money request, sexual request, friendly, geolocation). 3. For each of the above (input type, each label and stage), provide a concise rationale citing exact phrases or patterns in the message that justify your classification.” The output of the LLM results is in a customized format

Explanation and Prevention Advice Generation

For generating explanations of the scam stage, the GPT-4o model was adopted, and the prompt input for generating the reasons of the input texts and the scam stage are as follows: “I just detected a message, classified as stage 5 (Continuation of the Scam), the trigger factors are [trigger factors]. Please use 2 to 3 sentences to briefly explain why you made such a judgment” The prompt input for proving more information for users to understand of the scam stage is as follows: “I previously classified a message as stage 5 (Continuation of the Scam) with input type 'dialogue'. The detected triggers are: [trigger names]. Existing brief rationale: input type: dialogue; stage: 5; labels: [labels]. Please provide a more detailed explanation (4-5 sentences) of why this classification makes sense, including how the combination of signals supports the stage and any caveats or uncertainties” For providing prevention advice to reduce the risk of romance scam, the GPT-4o model was adopted, and the prompt input for generating prevention advice of the scam stage is as follows: “Given fraud stage 5 (Continuation of the Scam) and trigger factors [trigger names], provide exactly three practical prevention suggestions. Return them as a numbered list, each in one sentence” The prompt input for proving more prevention information of the scam stage for users is as follows: “Earlier you gave three concise prevention suggestions for fraud stage 5 (Continuation of the Scam) based on the trigger factors: [trigger names]: [summary list text]. Please expand each of the three suggestions into a detailed explanation. For each one, explain why it's important and how to implement it in practice, using 2 to 3 sentences per item”

Evaluation Results

Real romance scam data (Data number: 152) were collected from the Reddit website and were labeled with seven stages of the romance scam manually for evaluation From the detection results of the chatbot (Table 1), in identifying the stages of the romance scam have high accuracies in The sexual abuse (100%) and Continuation of the scam (92.9%) For “The sexual abuse” stage involves clear features such as sexual inducements or request of nude photos that the chatbot classified correctly, and the “Continuation of the scam” also have clear features such as urgent assistance or money request Wrong classification happens when the data have obvious features of other stages in the condition without context information, for example the data of “Will has sent Maria over 30,000 over the last seven years” in the “Continuation of the scam” stage was identified as The re-victimisation (re-victimization) stage The evaluation results show that the chatbot detects romance scam relies on context information in real scenario, and definitely performs better than the evaluation results

Conclusion

The chatbot provides a useful application for reducing the risk of romance scam, and can be applied in real scenario to provide real-time detection in Line chatting The explanations and the prevention advice of the detected scam stage also provide useful information to understand the stages of romance scam and to alert with any suspicious chatting on social media

Acknowledgments

This work was partially supported by the National Science and Technology Council, Taiwan, under grants 114-2635-E-004 -001-. Special thanks to the reporter, Bo-wei Hsu

References

[1] F. Wang, and V. Topalli, “Understanding romance scammers through the lens of their victims: Qualitative modeling of risk and protective factors in the online context,” Am J Crim Justice, vol. 49, pp. 145–181, Nov. 2024

[2] K.-F. Lee, M. Y. Chan, and A. Mohamad Ali, “Self and desired partner descriptions in the online romance scam: a linguistic analysis of scammer and general user profiles on online dating portals,” Crime Prev Community Saf, vol. 25, pp. 20–46, Mar. 2023

[3] A. Coluccia, A. Pozza, F. Ferretti, F. Carabellese, A. Masti, and G. Gualtieri, “Online romance scams: relational dynamics and psychological characteristics of the victims and scammers. A scoping review,” Clin Pract Epidemiol Ment Health, vol. 16, pp. 24–35, Mar.

Improvements for AI systems

  1. Enhance Contextual Analysis Precision: Improve context analysis by ensuring GPT-4o effectively distinguishes between pure dialogue and personal experience description to accurately determine if a message is authentic or indicative of a scam, as specified in the prompt: 1. Decide whether the message is a personal experience description or pure dialogue.

  2. Refine Stage Classification Mapping: Optimize the integer mapping for stage identification by ensuring clear differentiation between similar stages, such as stage 1 and stage 2, based on nuanced linguistic features extracted from user history to avoid misclassification errors highlighted in Table I.

  3. Improve Explanatory Depth via Trigger Analysis: Enhance the explanation module by utilizing the extracted trigger factors more deeply in the LLM output, ensuring the rationale explicitly links specific keywords (e.g., economic extraction, emotional manipulation) directly to the chosen stage, as shown in Figure 1 (a) and (b).

  4. Strengthen Prevention Advice Specificity: Upgrade the prevention module's output by making the three practical suggestions more granular and actionable, ensuring they are tailored precisely to the detected stage and its associated trigger factors to provide highly relevant risk reduction strategies.

  5. Implement Dynamic Visual Feedback System: Develop a dynamic visual feedback mechanism where the color coding for the seven stages is not static but is modulated based on real-time analysis of the input text's complexity (long vs. single sentence) to provide immediate, context-aware stage identification for users.

Related papers