Calibration of Machine Reading Systems at Scale

Findings of the Association for Computational Linguistics: ACL 2022, pages 1682 - 1693 May 22-27, 2022 c⃝2022 Association for Computational Linguistics Calibration of Machine Reading Systems at Scale Shehzaad Dhuliawala1, Leonard Adolphs1, Rajarshi Das2, Mrinmaya Sachan1 1ETH Zürich, 2University of Massachusetts Amherst {firstname.lastname}@inf.ethz.ch Abstract In typical machine learning systems, an es- timate of the probability of the prediction is used to assess the system’s confidence in the prediction. This confidence measure is usu- ally uncalibrated; i.e. the system’s confidence in the prediction does not match the true prob- ability of the predicted output. In this pa- per, we present an investigation into calibrat- ing open setting machine reading systems such as open-domain question answering and claim verification systems. We show that calibrat- ing such complex systems which contain a dis- crete retrieval and deep reading components is challenging and current calibration techniques fail to scale to these settings. We propose simple extensions to existing calibration ap- proaches that allow us to adapt these callibra- tors to these settings. Our experimental results reveal that the joint callibration of the retriever and the reader outperforms the reader calibra- tor by a significant margin. We also show that the callibrator can be useful for selective prediction, e.g., when question answering sys- tems are posed with unanswerable or out-of- the-training distribution questions. 1 Introduction With recent advances in machine reading, there has been a surge of interest in practical appli- cations of the technology such as open-domain question answering (Karpukhin et al., 2020; Lee et al., 2019) and claim verification (Thorne et al., 2018b). Due to various scale limitations in practi- cal settings, these systems are seldom trained end- to-end. Such systems typically make use of a RE- TRIEVER alongside a READER – the evidence is first retrieved from a large corpus and is then used by a machine reading model to provide an answer. As these systems are increasingly being de- ployed in the real world, it is important that they are not only accurate but also trustworthy. A way to make these systems trustworthy is to indicate when they are likely to be incorrect by provid- ing a calibrated confidence measure in addi- tion to the prediction. A naive solution for this is to use the system’s output probability as the confidence. However, this confidence score is of- ten uncalibrated (Kuleshov and Liang, 2015; Guo et al., 2017); i.e. it is not representative of the true correctness likelihood.1 Previous work (Jiang et al., 2020; Jagannatha and Yu, 2020; Desai and Durrett, 2020) has shown that large language models especially suffer from miscalibration. Thus, several methods have been proposed to calibrate language models based on gradient-based calibration methods such as tem- perature scaling (Guo et al., 2017) and feature- based forecasters (Kuleshov and Liang, 2015). While gradient-based calibration is intuitive and easy to implement, feature-based forecasters re- quire manual feature engineering. In this work, we contribute a simple method to calibrate practical RETRIEVER - READER ma- chine reading pipelines. These systems typically include …

Content truncated to 500 words