Neural networks, deep learning, NLP · insight-0454
Check whether a text model learned the template instead of the content
- Evidence type
- practitioner heuristic
- Difficulty
- medium
- Roles
- Data scientist, Applied AI
- Industries
- telecom, healthcare
Summary: situation, action, reason, boundary
Situation
A ticket classifier scores very high, and many tickets come from web forms, partner systems or agents with their own boilerplate.
Action
Slice errors and accuracy by source channel and template. Strip or mask boilerplate and retrain; if accuracy drops sharply, the model was using the template. Hold out a whole channel to test transfer.
Reason
Labels often correlate with how text was produced; the model uses any cue that predicts the label, whether or not it will persist.
Boundary
Ticket routing, document classification, fraud text and medical notes with templated sections.
Symptom or risk
The model keys on form headers, signatures or channel-specific phrases correlated with the label, then fails when a form changes or a new channel launches.
Exceptions
If the channel itself is a legitimate, stable signal, add it as an explicit feature instead.
Trade-offs
Removing boilerplate costs effort and may remove some genuine signal.
Pitfalls
Explaining the model with top n-grams and missing that they are signature lines.