An annotation artifact is an unintended pattern in evaluation data that correlates with a label or answer without reflecting the capability being tested. A system can exploit the pattern and score well while failing to perform the intended task. Artifacts can arise from annotation conventions, templated phrasing, label leakage, or correlations unrelated to the task. section 4.2.1 section 4.5.4
Hypothesis-only natural language inference
Natural language inference asks whether a hypothesis is entailed by, contradicts, or is neutral with respect to a premise. A model given only the hypothesis should perform at chance level if the label depends on the relationship between both sentences. Yet a hypothesis-only model reached about 67% accuracy on SNLI and over 50% on MultiNLI. Words such as “nobody,” “never,” and “no” were strongly associated with contradiction labels, while generic terms such as “animal” and “outdoors” tended to signal entailment. When systems were evaluated only on examples the hypothesis-only model could not solve, their performance dropped significantly. section 4.2.1
A strong benchmark score is therefore not enough to establish the intended capability. The key question is whether a recurring cue is genuinely informative for the decision or merely a by-product of data creation. section 4.2.1 section 4.5.4
To investigate suspected artifacts:
- Try a trivial or shallow baseline that should not succeed if the task requires the intended reasoning or understanding. Strong performance is a red flag. section 4.5.4
- Inspect examples the system gets right with very high confidence for repeated phrases, answer structures, verbosity, or other regularities in inputs, outputs, and labels. section 4.5.4
- Remove, mask, or neutralize suspected incidental cues and rerun the evaluation. A substantial performance drop is strong evidence that the original score was inflated by artifacts. section 4.5.4
