Ravindra Dastikop with ChatGPT
The college office has decided to try an AI system that sorts student messages. Two staff members agree on how to label the first few enquiries. Then they reach this one.
“I paid the fee yesterday, which was the last date, but the portal still says it is due.”
One staff member marks it Deadline because the message mentions the last date. The other marks it Payment status because the student wants to know whether the payment was recorded.
If the purpose is to route the message to the right office, Payment status is the more useful label. The deadline explains the urgency. The payment record is what someone needs to check.
Before a model can learn from examples, people need to decide what those examples should teach.
2.1 What counts as data?
Data is recorded information that a computer can process. A student message, a photograph, and a row in a spreadsheet are all examples of data.
For the enquiry sorter, one example contains a student’s message and the category a staff member assigns to it. A collection of these examples is a dataset.
| Student message | Label |
|---|---|
| “What is the last date to pay the examination fee?” | Deadline |
| “The money has left my bank account, but the portal says it is due.” | Payment status |
| “I missed the deadline. Can I still pay?” | Late payment |
The message is the input. The assigned category is the label. When a model learns from inputs paired with labels, the approach is called supervised learning.
The sorter can work with the wording of an enquiry. Confirming whether a payment succeeded requires the actual payment record. The college can keep those private records out of the sorter’s training dataset.
2.2 Write a label guide
A label guide explains what each category means and helps staff handle unclear messages consistently.
| Label | What the student wants to know |
|---|---|
| Deadline | The scheduled date by which payment is due. |
| Payment status | Whether a payment they already attempted was recorded. |
| Late payment | Whether they can pay after the deadline or get an exception. |
“The deadline was yesterday. I paid on time, but the portal hasn’t updated.”
The student wants to know whether a payment already made was recorded. Under this guide, the label is Payment status.
“My payment failed. Can I try again tomorrow even though the deadline has passed?”
If the college needs one label for routing, it could tell staff to follow the student’s explicit question. That makes this Late payment. If the college needs to track both issues, it could permit multiple labels.
Ask two staff members to label the same small batch independently. Where they disagree, ask what each understood the student to be requesting. Use those disagreements to clarify the guide before labelling a larger batch.
2.3 Collect examples of how people actually ask
A hundred messages that all say “What is the last date?” give the model a hundred examples of nearly the same wording. Students may also ask:
- “How much time do we have to pay?”
- “When does the fee window close?”
- “Is Friday the final day for exam fee payment?”
All three ask about a deadline. Include varied wording, and if students commonly use several languages or mix them in one message, include that variety and test the system on it.
Include difficult cases too. “I paid after the last date, but the portal shows nothing” could be a payment-status question. More deadline examples may do little to help if the model keeps confusing late-payment requests with deadline enquiries.
2.4 Keep some examples for a fair test
Imagine a student gets every question right on an exam that repeats the practice sheet. You still need to see how the student handles a new question.
A training set contains examples the model learns from. A test set contains separate examples used to check how the trained model handles new inputs. If a student sends the same enquiry by email and through the portal, keep both copies on the same side of the split. Otherwise, the model could get credit for correctly handling a message it effectively saw during training.
Check results by category. Suppose a model gets all 18 deadline questions right and both late-payment questions wrong. Its overall accuracy is 18 out of 20, or 90%. On late-payment questions it scores 0 out of 2. Two examples are too few to estimate its failure rate for the whole category, but those failures tell you where to test further.
2.5 Try it yourself · Label the college enquiries
Use the three labels from Section 2.2. Mark a message Needs review if it gives you too little information to route it. This flag tells a person to look at the message.
- “When does the exam fee window close?”
- “My bank shows the amount was debited, but I still don’t have a receipt.”
- “I missed the final date. Is there another chance to pay?”
- “I paid before the last date, but the portal still shows the amount as due.”
- “I have a problem with the exam fee. Please help.”
- “Is tomorrow the final day to pay?”
Assign a label or review flag to each message. Underline the words that tell you what the student needs. Write one sentence explaining how you labelled message 4. Then set aside messages 4 and 6 as practice test cases. If a model labelled both Deadline, which prediction would you correct?
Check your work
Messages 1 and 6 are Deadline. Messages 2 and 4 are Payment status. Message 3 is Late payment. Message 5 Needs review. A useful follow-up would ask whether the student needs the due date, help with a payment already attempted, or permission to pay late.
Correct message 4. The student mentions the last date but wants to know why an earlier payment still shows as due. You have a small labelled dataset and a test case that exposes the weakness of relying on a keyword. A working sorter would need a broader sample.
Chapter summary
- In supervised learning, the examples pair inputs with targets such as category labels.
- Labels reflect the task the system is meant to perform. A written guide helps people apply them consistently.
- Include varied wording and difficult cases. Leave out personal details the task does not require.
- Test on separate examples and examine errors by category. An overall score can hide weaknesses that matter to users.
Review questions
- What is the input and what is the label in a training example for the enquiry sorter?
- Why might two staff members label the same message differently? How could a label guide help?
- Why should copies of the same enquiry stay on the same side of the training and test split?
- A sorter works well on deadline questions and often misses late-payment requests. What examples would you collect next?
- Write a message that mentions a deadline but should be labelled Payment status. Explain why.
Next chapter · How a Model Learns
How does a model turn labelled examples into a prediction about a message it has never seen?

.png)

.png)


