Do Explanations Help Users Detect Errors in Open-Domain QA? An Evaluation of Spoken vs. Visual Explanations

Findings of the Association for Computational Linguistics: ACL-IJCNLP 2021, pages 1103–1116 August 1–6, 2021. ©2021 Association for Computational Linguistics 1103 Do Explanations Help Users Detect Errors in Open-Domain QA? An Evaluation of Spoken vs. Visual Explanations Ana Valeria González∗1, Gagan Bansal2, Angela Fan3,4, Yashar Mehdad3, Robin Jia3, and Srinivasan Iyer3 1University of Copenhagen, 2University of Washington, 3Facebook AI, 4LORIA ana@di.ku.dk bansalg@cs.washington.edu {angelafan, mehdad, robinjia, sviyer}@fb.com Abstract While research on explaining predictions of open-domain QA systems (ODQA) is gain- ing momentum, most works do not evaluate whether these explanations improve user trust. Furthermore, many users interact with ODQA using voice-assistants, yet prior works exclu- sively focus on visual displays, risking (as we also show) incorrectly extrapolating the effec- tiveness of explanations across modalities. To better understand the effectiveness of ODQA explanations strategies in the wild, we con- duct user studies that measure whether expla- nations help users correctly decide when to accept or reject an ODQA system’s answer. Unlike prior work, we control for explana- tion modality, i.e., whether they are communi- cated to users through a spoken or visual inter- face, and contrast effectiveness across modali- ties. We show that explanations derived from retrieved evidence can outperform strong base- lines across modalities but the best explanation strategy varies with the modality. We show common failure cases of current explanations, emphasize end-to-end evaluation of explana- tions, and caution against evaluating them in proxy modalities that differ from deployment. 1 Introduction Despite copious interest in developing explainable AI, there is increasing skepticism as to whether explanations (of system predictions) are useful to end-users in downstream applications. For instance, for assisting users with classifying sentiment or an- swering LSAT questions, Bansal et al. (2021) ob- served no improvements from giving explanations over simply presenting model confidence. Simi- larly, Chu et al. (2020) observed that visual expla- nations fail to significantly improve user accuracy or trust. Such negative results present a cautionary tale for explainability and emphasize the need to evaluate explanations using careful user studies. ∗ Work done while at Facebook AI. Figure 1: Using end-to-end user studies, we evaluate whether explanation strategies of open-domain QA as- sistants help users decide when to trust (or reject) pre- dicted answers. We explore the effectiveness of explanations for Open-Domain Question Answering models, which involves answering users’ questions (e.g., “Who plays the Joker in the Lego Batman movie?”) using a large corpus (e.g., Wikipedia). Such models are increasingly deployed not only in visual modalities (e.g., Web search) but also in spoken ones (voice assistants).1 Spoken interfaces for ODQA are also important because they make ODQA systems more accessible for users with visual impairments. De- spite improvements in accuracy, deployed ODQA models remain imperfect. This motivates the need to provide users with mechanisms (e.g., estimates of uncertainty or explanations) that can help im- prove appropriate reliance (Lee and See, 2004), e.g., by allowing users to detect erroneous answers. We henceforth refer to a user’s ability to distinguish correct and incorrect answers as error-detectability, and ask Does explaining …

Content truncated to 500 words