Misleading Failures of Partial-input Baselines

Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, pages 5533–5538 Florence, Italy, July 28 - August 2, 2019. c⃝2019 Association for Computational Linguistics 5533 Misleading Failures of Partial-input Baselines Shi Feng Computer Science University of Maryland shifeng@umiacs.umd.edu Eric Wallace Allen Institute for Artificial Intelligence ericw@allenai.org Jordan Boyd-Graber Computer Science, iSchool, UMIACS, and LSC University of Maryland jbg@umiacs.umd.edu Abstract Recent work establishes dataset difficulty and removes annotation artifacts via partial-input baselines (e.g., hypothesis-only models for SNLI or question-only models for VQA). When a partial-input baseline gets high accuracy, a dataset is cheatable. However, the converse is not necessarily true: the failure of a partial- input baseline does not mean a dataset is free of artifacts. To illustrate this, we first design ar- tificial datasets which contain trivial patterns in the full input that are undetectable by any partial-input model. Next, we identify such ar- tifacts in the SNLI dataset—a hypothesis-only model augmented with trivial patterns in the premise can solve 15% of the examples that are previously considered “hard”. Our work provides a caveat for the use of partial-input baselines for dataset verification and creation. 1 Dataset Artifacts Hurt Generalizability Dataset quality is crucial for the development and evaluation of machine learning models. Large- scale natural language processing (NLP) datasets often use human annotations on web-crawled data, which can introduce artifacts. For example, crowd- workers might use specific words to contradict a given premise (Gururangan et al., 2018). These ar- tifacts corrupt the intention of the datasets to train and evaluate models for natural language under- standing. Importantly, a human inspection of indi- vidual examples cannot catch artifacts because they are only visible in aggregate on the dataset level. However, machine learning algorithms, which de- tect and exploit recurring patterns in large datasets by design, can just as easily use artifacts as real lin- guistic clues. As a result, models trained on these datasets can achieve high test accuracy by exploit- ing artifacts but fail to generalize, e.g., they fail under adversarial evaluation (Jia and Liang, 2017; Ribeiro et al., 2018). The identification of dataset artifacts has changed model evaluation and dataset construc- tion (Chen et al., 2016; Jia and Liang, 2017; Goyal et al., 2017). One key method is to use partial- input baselines, i.e., models that intentionally ig- nore portions of the input. Example use cases in- clude hypothesis-only models for natural language inference (Gururangan et al., 2018), question-only models for visual question answering (Goyal et al., 2017), and paragraph-only models for reading com- prehension (Kaushik and Lipton, 2018). A success- ful partial-input baseline indicates that a dataset contains artifacts which make it easier than ex- pected. On the other hand, examples where this baseline fails are “hard” (Gururangan et al., 2018), and the failure of partial-input baselines is consid- ered a verdict of a dataset’s difficulty (Zellers et al., 2018; Kaushik and Lipton, 2018). These partial-input analyses are valuable and in- deed reveal dataset issues; however, they do not tell the whole story. Just …

Content truncated to 500 words