On Evaluating Explanation Utility for Human-AI Decision-Making in NLP
Fall 2023 XAIA NeurIPS 2023 Workshop

We can use imperfect AI models effectively if we can determine when they are correct. To achieve this, providing explanations has been suggested, hoping that these explanations can illuminate the reasoning process of the AI models. Past work in the explainability field primarily concentrated on technical challenges and proxy evaluations, leading to a gap in human-centered evaluations of explanations. As explainability methods advance beyond the proof-of-concept stage, our work aims to offer guidelines for navigating this evolving landscape, drawing upon over 50 existing datasets. We assess the suitability of these datasets for application-grounded evaluation of the explanations’ usefulness in real-world high-stakes situations. Furthermore, we emphasize the importance of adapting task setups to reflect realistic scenarios (such as incorporating text retrieval components), reevaluating the performance of backbone models in the fast-evolving NLP field, and refining experimental protocols to accurately measure the impact of explanations on human reliance and human-AI team performance. Our work provides a practical approach for future research in human-centered evaluation of the efficacy of various explanations in real-world situations. Combining the gathered insights, analyses, and developed models, the next phase is evaluating the explanations for the four datasets that are proven to be suitable through our meta-analysis.

Salek Faramarzi, N.^*, Hashemi Chaleshtori, Fateme^*, Shirazi, H., Banerjee, R., & Ray, I. (Apr. 2023). ”Claim Extraction and Dynamic Stance Detection in COVID-19 Tweets”. Companion Proceedings of the ACM Web Conference. (*equal contribution)
Link to paper

Hashemi Chaleshtori, Fateme & Ray, I. (Dec. 2022). ”Automation of Vulnerability Information Extraction Using Transformer-Based Models of Natural Language Processing”. European Symposium on Research in Computer Security (ESORICS 2022) Workshop on System Security Assurance (SecAssure 2022).
Link to paper

Zou, C., Banerjee R., Hashemi Chaleshtori, Fateme Shirazi, H., & Ray, I. (Aug. 2021). ”Seeing Should Probably not be Believing: The Role of Deceptive Support in COVID-19 Misinformation on Twitter”. . Journal of Data and Information Quality
Link to paper

Rahimi, M., Shirazi, M., NajafGholian, A., Hashemi Chaleshtori, Fateme, Moradi, N., Behzad, K., Roodabeh, S. H., Gavahi, A., Moghadam, F. F., GhaziAsgar, A., AlizadehGharib, Y., Memarian, M., Tavakoli, H. & Khosravi, M. A. ”Parsian 2018 Extended Team Description Paper for RoboCup”. International RoboCup Symposium.
Link to report

Rahimi, M., Shirazi, M., Arfaee, M., NajafGholian, A., Zamani, A., Hosseini, H., Hashemi Chaleshtori, Fateme, Moradi, N., Ahsani, A., Jafari, M., Zahedi, A., Abdollahi, P. & Khosravi, M. A. ”Parsian 2017 Extended Team Description Paper for RoboCup”. International RoboCup Symposium.
Link to report