先自然见面，再把研究讲清楚。

Let me start with the practical problem. I studied Computer Science and Technology, and I now work full-time in Sydney in the online-trading industry. My work experience made me notice that a promoter, a victim, and a person warning others may repeat very similar claims in Chinese trading communities. If a system only sees phrases such as “guaranteed profit” or “deposit now”, it may confuse these very different speakers. My goal is therefore not to let AI decide who is a scammer. It is to rank messages more carefully, so that a human reviewer can look at the most relevant items first.

My main question is this: under the same number of training labels, can a case-informed, auditable and stance-aware weak-supervision method rank risky promotion better than a standard human-labelled supervised baseline on an unseen channel, without clearly increasing harmful false alerts on victim accounts and warnings? The main comparison is not between large language models. It is between two ways of creating training signals, tested in a setting that is closer to real use.

First, I need to define the task clearly. For each risk claim, I separate three things. Conduct means what appears in the message, such as a guaranteed-return claim or pressure to deposit. Stance means what the speaker is doing: promoting the claim, describing personal harm, quoting it, or warning other people. Outside evidence means information that can be checked separately, such as a complaint, a regulatory warning or a court record. For example, a victim may say, “The group promised stable profit, but I lost money after depositing.” The risky words are present, but the speaker is reporting harm, not promoting the claim. Outside evidence remains separate and does not replace the text label.

Second, I compare two training routes. Route A uses permitted messages labelled by people under the same annotation guide. This creates a human-labelled training set. Route B turns clear decisions from public official cases into written, auditable rules. A rule may vote positive, vote negative, or abstain. These votes are combined into uncertain weak labels, which form a weak-labelled training set. I match the number of training labels and use exactly the same learning recipe twice. The output is two trained models that can each give a risk score to a new message.

For a fair first comparison, both routes use the same simple learner. TF-IDF turns words and short phrases into numeric features. It is like translating text into a table that a computer can calculate. Logistic regression then learns the link between those features and the training labels, and gives each new message a risk score. Route A trains model A, and route B trains model B. At this point both models can answer questions, but we still do not know which one learned better. Training performance is not the final exam result.

The next stage is the exam. Written permission comes first. Before training, I reserve one complete channel for testing. Repeated or near-duplicate reposts are grouped, so the same message family cannot appear on both sides of the training and test boundary. After training, both models are frozen. They then score the same test messages, which they have never used for training. A separate group of human annotators prepares the gold set, or human answer key, for those messages. The models give risk scores to messages; the evaluation software compares those scores with the answer key and calculates a report card for each model. These are two different steps: the student answers, and the teacher marks the exam.

The primary measure is Average Precision, or AP. In simple terms, it asks whether the messages that truly need review stay near the top of the full ranking. I also compare the models under the same human-review budget, for example the top fifty messages per day. Safety is checked separately on difficult negative cases, especially victim accounts, warnings and ordinary trading discussion. I set the acceptable limit before seeing the test results. If possible, I would later keep the models frozen and repeat the offline test on a later approved time window. “Future” here means a later collected test batch, not live monitoring.

The system only produces a ranked queue for human review. It does not decide that a person, company or channel is illegal, and it does not replace regulatory, investigative or legal judgement. The research asks which training signal makes better use of limited human attention when labels are expensive, language changes across channels, and harmful mistakes matter.

At present, I have prepared a public research protocol and a reproducibility companion. The research question, comparison, test boundary and stopping conditions are written down. However, I do not yet have a permitted multi-channel Chinese message stream, formal Chinese labels, trained Chinese models or experimental results. I am not presenting the method as already proven. The next step should be a governance and ethics feasibility check, followed by a small double-coded pilot using public official cases to test whether the task and rules can be applied consistently.

I wanted to discuss this project with you because the core problem is not financial-market prediction. It is reliable and explainable machine-learning evaluation when data changes, speaker stance is ambiguous, and errors can cause real harm. I see a natural connection with your work on reliable machine learning, misinformation and trustworthy AI. For an MPhil project, I would especially value your view on whether the scope is realistic, what should be reduced first, what parts of the method need redesign, and whether the direction could fit your supervision or a co-supervision arrangement.