Tekst measures how well a matching model is performing by comparing the record it picks against a test set of validated examples. This article explains how that score is calculated and how to improve it.
If you have not built a test set yet, see Set up your first matching model first.
Exact-match scoring
Matching uses a simple, strict rule: for each test item, the model's answer is correct only if it picks the exact record you confirmed as the ground truth, and wrong otherwise. There is no partial credit - matching to a similar but different record counts as wrong, because in practice the wrong record is the wrong answer.
The accuracy you see is the number of correct matches divided by the total number of test items.
Reviewing wrong matches
For each test item, Tekst shows the record the model picked next to the record you confirmed, along with the reasoning the model gave for its choice. Comparing the two is the fastest way to understand a wrong answer: the reasoning usually shows whether the model misread the input, or read it correctly but had no good record to pick.
When accuracy cannot be calculated
Accuracy needs both imported records and test items with confirmed records. If you have not imported records yet, or have not added test items, the model will tell you what is missing before it can report a score. If a re-import removes a record that a test item referred to, review the flagged item before relying on the figure.
Improving the score
There are two levers, and the wrong matches tell you which one to reach for:
- Fix your records when the right answer is not in your data, or is too sparsely described to be recognizable. Missing entries and thin descriptions are a common cause of wrong matches.
- Refine your domain knowledge when the right record exists but the model does not connect the wording to it. Synonyms, abbreviations, and tie-breaking rules belong here.
As your team confirms and corrects matches, those corrections also feed back into the model.
Related articles
- To maintain the records the model matches against, see Managing master data records.
- To steer the model's choices, see Using domain knowledge to guide matching.
0 comments
Please sign in to leave a comment.