Baxnet Ideas · Technical explainer
Before Training a Model, Define the Topic
TL;DR: A classifier cannot repair a vague category. Before it can learn, someone has to decide what the label means.
It is tempting to start a topic experiment with a spreadsheet and a model button. We have a label called “music”; surely the first task is to collect examples of people saying “album”, “gig”, and “playlist”.
The harder work comes first. What does music mean for this label?
Consider these fictional messages:
“I finally learned the first eight bars.”
“The builders said the music room needs more insulation.”
“That song came on during the scan and I hated it.”
“Can you send the music for the slideshow?”
All four contain a plausible musical reference. They do not necessarily belong in the same category for every product question. The first could be about practising an instrument. The second is construction. The third might be healthcare context. The fourth is planning an event. A good topic definition says what it wants to capture, what it excludes, and where a separate label would be more honest.
This is a meaning decision before it is a machine-learning decision. If two careful people cannot explain why a message belongs in or out of a topic, a training run will only turn that uncertainty into a cleaner-looking number.
Start with the useful boundary
The most useful definition is usually narrower than the label’s everyday name. It is built around a question a person might actually ask.
For a topic called “home maintenance”, a workable scope might include arranging repairs, discussing faults, comparing tradespeople, and planning routine maintenance. It might exclude home decoration, real-estate speculation, and a vague line such as “this place is falling apart” when the context is a difficult week rather than a ceiling leak.
Those exclusions are not failures. They make the label less likely to become a grab bag for any message that shares a word with a house.
The boundary also needs difficult negatives: examples that look close but should remain out. “The doctor fixed it” should not teach a repairs topic. “I need to rebuild my routine” is not evidence of construction. “The account needs maintenance” may belong to work administration. These are the examples that expose whether a label follows meaning or merely familiar vocabulary.
A gazetteer is evidence, not truth
One helpful tool is a gazetteer: a governed list of words and phrases that can point an experiment toward evidence. In a home-maintenance topic, that list might include “boiler service”, “leaking tap”, “electrician”, or a regionally specific name for a common repair.
The list can be valuable. It gives reviewers a visible starting point and helps us find cases worth checking. It should not become the definition of the topic.
Words travel. “Leak” appears in plumbing, software, gossip, and medical conversations. “Service” can mean a car appointment, a church gathering, a restaurant, or customer support. A phrase list can find candidates; it cannot decide every case without context. Calling it governed evidence is a useful discipline because it keeps the list open to review and change instead of treating it as a miniature dictionary of human truth.
Practical home maintenance.
Different subject.
Insufficient context.
The same humility applies to model outputs. A model may see a relationship between phrases that a list misses. That can be useful, but it still has to be checked against the adopted boundary. A clever prediction outside the boundary is still wrong for that experiment.
Examples make the policy real
Once a boundary exists, examples become less like raw material and more like test cases for a policy.
We want ordinary positive cases, of course. We also want near misses, brief references, ambiguous shorthand, and examples from different ways of speaking. A topic that only learns from tidy sentences will be surprised by the way people actually write: half a thought, a typo, a photo caption, a response to a message we cannot see, or a subject that changes halfway through a sentence.
This does not mean collecting every possible variant until the definition disappears. It means recording why a case was admitted or excluded, then returning to the definition when reviewers disagree. Sometimes the right outcome is a refinement. Sometimes it is an unknown case that should not be forced into either pile.
That process is slow in an unglamorous way. It is also where the quality comes from. A model trained on a crisp definition can still make mistakes. A model trained on a shifting definition is difficult to evaluate consistently, because the target changes whenever a result feels inconvenient.
The decision remains visible
In Mimoto’s experimental work, the aim is to keep the definition, examples, split rules, and evaluation receipts inspectable. That makes it possible to ask a useful question after a result arrives: did the model misunderstand the topic, or did we discover that our definition was not ready?
That question is much more productive than asking whether the model is “smart”. A model does not own the category. The people building and reviewing the experiment do.
The next piece looks at a particular family of tools that may help with this work: text embeddings and transfer learning. They can recognise more than exact phrases, but they do not remove the need for the boundary we have just drawn.