Selective Question Answering under Domain Shift

Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 5684–5696 July 5 - 10, 2020. c⃝2020 Association for Computational Linguistics 5684 Selective Question Answering under Domain Shift Amita Kamath Robin Jia Percy Liang Computer Science Department, Stanford University {kamatha, robinjia, pliang}@cs.stanford.edu Abstract To avoid giving wrong answers, question an- swering (QA) models need to know when to abstain from answering. Moreover, users of- ten ask questions that diverge from the model’s training data, making errors more likely and thus abstention more critical. In this work, we propose the setting of selective question an- swering under domain shift, in which a QA model is tested on a mixture of in-domain and out-of-domain data, and must answer (i.e., not abstain on) as many questions as possi- ble while maintaining high accuracy. Ab- stention policies based solely on the model’s softmax probabilities fare poorly, since mod- els are overconfident on out-of-domain inputs. Instead, we train a calibrator to identify in- puts on which the QA model errs, and ab- stain when it predicts an error is likely. Cru- cially, the calibrator benefits from observing the model’s behavior on out-of-domain data, even if from a different domain than the test data. We combine this method with a SQuAD- trained QA model and evaluate on mixtures of SQuAD and five other QA datasets. Our method answers 56% of questions while main- taining 80% accuracy; in contrast, directly using the model’s probabilities only answers 48% at 80% accuracy. 1 Introduction Question answering (QA) models have achieved impressive performance when trained and tested on examples from the same dataset, but tend to per- form poorly on examples that are out-of-domain (OOD) (Jia and Liang, 2017; Chen et al., 2017; Yogatama et al., 2019; Talmor and Berant, 2019; Fisch et al., 2019). Deployed QA systems in search engines and personal assistants need to gracefully handle OOD inputs, as users often ask questions that fall outside of the system’s training distribution. While the ideal system would correctly answer all Dataset Distributions Example question Q: What can result from disorders of the immune system? (from SQuAD) Q: John Wickham Legg was recommended by Jenner for the post of medical attendant to which eighth child and youngest son of Queen Victoria and Prince Albert of Saxe-Coburg and Gotha? (from HotpotQA) Q: Capote gained fame with this “other” worldly 1948 novel about a teenager in a crumbling southern mansion. (from SearchQA) Train Calibrate Test Source Source Known OOD Source Unknown OOD Figure 1: Selective question answering under domain shift with a trained calibrator. First, a QA model is trained only on source data. Then, a calibrator is trained to predict whether the QA model was correct on any given example. The calibrator’s training data con- sists of both previously held-out source data and known OOD data. Finally, the combined selective QA system is tested on a mixture of test data from the source dis- tribution and an unknown OOD distribution. OOD questions, such perfection is not attainable given …

Content truncated to 500 words