We Need to Talk About Random Splits
This addresses a methodological problem for NLP researchers by highlighting flaws in common evaluation practices, suggesting more realistic benchmarks.
The paper argues that both random and standard data splits in NLP experiments produce overly optimistic performance estimates, and shows that even worst-case biased splits often underestimate error on new in-domain data, invalidating the covariate shift assumption.
Gorman and Bedrick (2019) argued for using random splits rather than standard splits in NLP experiments. We argue that random splits, like standard splits, lead to overly optimistic performance estimates. We can also split data in biased or adversarial ways, e.g., training on short sentences and evaluating on long ones. Biased sampling has been used in domain adaptation to simulate real-world drift; this is known as the covariate shift assumption. In NLP, however, even worst-case splits, maximizing bias, often under-estimate the error observed on new samples of in-domain data, i.e., the data that models should minimally generalize to at test time. This invalidates the covariate shift assumption. Instead of using multiple random splits, future benchmarks should ideally include multiple, independent test sets instead; if infeasible, we argue that multiple biased splits leads to more realistic performance estimates than multiple random splits.