We Need To Talk About Random Splits

We Need to Talk About Random Splits Anders Søgaard1,2 Sebastian Ebert1 Jasmijn Bastings1 Katja Filippova1 1: Google Research, Berlin 2: University of Copenhagen {soegaard|eberts|bastings|katjaf}@google.com Abstract Gorman and Bedrick (2019) argued for us- ing random splits rather than standard splits in NLP experiments. We argue that random splits, like standard splits, lead to overly op- timistic 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, maxi- mizing bias, often under-estimate the error ob- served on new samples of in-domain data, i.e., the data that models should minimally gener- alize to at test time. This invalidates the co- variate shift assumption. Instead of using mul- tiple random splits, future benchmarks should ideally include multiple, independent test sets instead; if infeasible, we argue that multiple bi- ased splits leads to more realistic performance estimates than multiple random splits. 1 Introduction It is common practice in NLP to collect and anno- tate a text corpus – and split it into training, de- velopment and test data. These splits are often based on the order in which texts were published or sampled, and are referred to as ‘standard splits’. Gorman and Bedrick (2019) recently showed that system ranking results based on standard splits dif- fer from results based on random splits and used this to argue in favor of using random splits. While perhaps less common, random splits are already used in probing (Elazar and Goldberg, 2018), in- terpretability (P¨orner et al., 2018), as well as core NLP tasks (Yu et al., 2019; Geva et al., 2019).1 Gorman and Bedrick (2019) focus on whether there is a significant performance difference δ 1See also many of the tasks in the SemEval evaluation campaigns: http://alt.qcri.org/semeval2020/ Figure 1: Data splitting strategies. Each ball corre- sponds to a sentence represented in (two-dimensional) feature space. Blue (dark)/orange (bright) balls repre- sent examples for training/test. Numbers represent sen- tence length. Heuristic splits can, e.g., be based on sen- tence length; adversarial splits maximize divergence. between systems S1 and S2; M(Gtest, S1) − M(Gtest, S2), in their notation. They argue Mc- Nemar’s test (Gillick and Cox, 1989) or bootstrap (Efron, 1981) can establish that δ ̸= 0, using ran- dom splits to sample from Gtest. This, of course, relies on the assumption that data is representative, i.e., was sampled i.i.d. (Wolpert, 1996). In reality, what Gorman and Bedrick (2019) call the true difference in system performance, i.e., δ = M(Gtest, S1) − M(Gtest, S2), is the system difference on data that users would expect the sys- tems to work well on (see §2 for practical examples) – and not just on the corpus that we have annota- tions for. Our corpus-based estimates of δ can in fact be very misleading, i.e., very different from performance on new samples of data. In this pa- …

Content truncated to 500 words