Neural networks, deep learning, NLP · insight-0460
Audit test-set labels before ranking models that differ by a point or two
- Evidence type
- empirical finding
- Difficulty
- medium
- Roles
- Data scientist, Applied AI
- Industries
- e-commerce
Summary: situation, action, reason, boundary
Situation
Two text classifiers differ by 1.5 points on a test set labelled quickly by several annotators.
Action
Double-annotate a sample of the test set, measure disagreement, review examples where the models and labels disagree most, and recompute the comparison on a cleaned subset.
Reason
Northcutt et al. estimated an average of at least 3.3% label errors in test sets of common benchmarks and found lower-capacity models could be preferable when test labels are noisy.
Boundary
Any evaluation where labels come from crowdsourcing, agents' tags or weak rules.
Symptom or risk
If several percent of test labels are wrong, the ranking can flip on a corrected set, and the more flexible model may be rewarded for fitting annotation quirks.
Exceptions
Labels produced by a reliable system of record, such as a ledger outcome, need less auditing.
Trade-offs
Annotation costs money and time; focus on disagreement-ranked samples.
Pitfalls
Cleaning only examples the current model gets wrong, which biases the test set in its favour.
Lessons and concepts
Claims and their sources
Northcutt et al. estimate an average of at least 3.3% label errors across test sets of 10 common datasets and find lower-capacity models may be more useful when test labels are noisy.