Validity, Reliability, and Significance: Empirical Methods for NLP and Data Science

When we come up with a new model in NLP and machine learning more generally, we usually look at some performance metric (one number), compare it against the same performance metric for a strong baseline model (one number), and if the new model gets a better number, we mark it in bold and declare it the winner. For anyone with a background in statistics or a field where conclusions must be drawn on the basis of noisy data, this procedure is frankly shocking. Suppose model A gets a BLEU score one point higher than model B: Is that difference reliable? If you used a slightly different dataset for training and evaluation, would that one point difference still hold? Would the difference even survive running the same models on the same datasets but with different random seeds? In fields such as psychology and biology, it is standard to answer such questions using standardized statistical procedures to make sure that differences of interest are larger than some quantification of measurement noise. Making a claim based on a bare difference of two numbers is unthinkable. Yet statistical procedures remain rare in the evaluation of NLP models, whose performance metrics are arguably just as noisy. To these objections, NLP practitioners can respond that they have faithfully followed the hallowed train-(dev-)test split paradigm. As long as proper test set discipline has been followed, the theory goes, the evaluation is secure: By testing on held-out data, we can be sure that our models are performing well in a way that is independent of random accidents of the training data, and by testing on that data only once, we guard against making claims based on differences that would not replicate if we ran the models again. But does the train-test split paradigm really guard against all problems of validity and reliability? Into this situation comes the book under review, Validity, Reliability, and Significance: Empirical Methods for NLP and Data Science, by Stefan Riezler and Michael Hagmann. The authors argue that the train-test split paradigm does not in fact insulate NLP from problems relating to the validity and reliability of its models, their features, and their performance metrics. They present numerous case studies to prove their point, and advocate and teach standard statistical methods as the solution, with rich examples