DARE: Towards Robust Text Explanations in Biomedical and Healthcare Applications

Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics Volume 1: Long Papers, pages 11499–11533 July 9-14, 2023 ©2023 Association for Computational Linguistics DARE: Towards Robust Text Explanations in Biomedical and Healthcare Applications Adam Ivankay IBM Research Zurich Rüschlikon, Switzerland aiv@zurich.ibm.com Mattia Rigotti IBM Research Zurich Rüschlikon, Switzerland mrg@zurich.ibm.com Pascal Frossard École Polytechnique Fédérale de Lausanne (EPFL) Lausanne, Switzerland pascal.frossard@epfl.ch Abstract Along with the successful deployment of deep neural networks in several application domains, the need to unravel the black-box nature of these networks has seen a significant increase recently. Several methods have been introduced to provide insight into the inference process of deep neural networks. However, most of these explainability methods have been shown to be brittle in the face of adversarial perturbations of their inputs in the image and generic tex- tual domain. In this work we show that this phenomenon extends to specific and important high stakes domains like biomedical datasets. In particular, we observe that the robustness of explanations should be characterized in terms of the accuracy of the explanation in linking a model’s inputs and its decisions – faithfulness – and its relevance from the perspective of do- main experts – plausibility. This is crucial to prevent explanations that are inaccurate but still look convincing in the context of the domain at hand. To this end, we show how to adapt cur- rent attribution robustness estimation methods to a given domain, so as to take into account domain-specific plausibility. This results in our DOMAINADAPTIVEARESTIMATOR (DARE) attribution robustness estimator, allowing us to properly characterize the domain-specific robustness of faithful explanations. Next, we provide two methods, adversarial training and FAR training, to mitigate the brittleness charac- terized by DARE, allowing us to train networks that display robust attributions. Finally, we em- pirically validate our methods with extensive experiments on three established biomedical benchmarks. 1 Introduction Research in explainable AI (XAI) has seen a surge in recent years. XAI methods aim to provide in- sight into the inference process and the causal links between inputs and outputs of deep neural networks (DNNs). This is pivotal in addressing many aspects of DNNs, such as fairness, potential biases and scopes of safe deployment. Especially in safety- critical domains, such as healthcare, faithful and robust explanations (Jacovi and Goldberg, 2020) accompanying the predictions of DNNs are key to enable their deployment and understand poten- tial false predictions and risks. Beside faithfulness, which quantifies the accuracy by which the explana- tions characterize the true decision-making process of the model, a second property of explanations that has been highlighted as important is plausi- bility (Jacovi and Goldberg, 2021; Rigotti et al., 2022; Rizzo et al., 2022). Plausibility quantifies the “degree to which some explanation is aligned with the user’s understanding” (Jacovi and Gold- berg, 2021). As such, plausibility tells us whether an explanation is found convincing and informa- tive for domain experts. These independent but related properties are therefore crucial ingredients to provide explanations that are accurate and robust, as …

Content truncated to 500 words