Modeling Disclosive Transparency in NLP Application Descriptions

Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 2023–2037 November 7–11, 2021. c⃝2021 Association for Computational Linguistics 2023 Modeling Disclosive Transparency in NLP Application Descriptions Michael Saxon, Sharon Levy, Xinyi Wang, Alon Albalak, William Yang Wang University of California, Santa Barbara {saxon, xinyi_wang, alon_albalak}@ucsb.edu, {sharonlevy, william}@cs.ucsb.edu Abstract Broader disclosive transparency—truth and clarity in communication regarding the func- tion of AI systems—is widely considered de- sirable. Unfortunately, it is a nebulous con- cept, difficult to both define and quantify. This is problematic, as previous work has demon- strated possible trade-offs and negative conse- quences to disclosive transparency, such as a confusion effect, where “too much informa- tion” clouds a reader’s understanding of what a system description means. Disclosive trans- parency’s subjective nature has rendered deep study into these problems and their remedies difficult. To improve this state of affairs, We introduce neural language model-based prob- abilistic metrics to directly model disclosive transparency, and demonstrate that they cor- relate with user and expert opinions of sys- tem transparency, making them a valid objec- tive proxy. Finally, we demonstrate the use of these metrics in a pilot study quantifying the relationships between transparency, confusion, and user perceptions in a corpus of real NLP system descriptions. 1 Introduction Among the draft Ethics Guidelines for Trustworthy AI released by the European Union in 2019 were calls for greater transparency around deployed sys- tems using artificial intelligence, advising that an “AI system’s capabilities and limitations should be communicated to practitioners or end-users in a manner appropriate to the use case at hand.” (High- Level Expert Group on AI, 2019) This is a high- profile example of the vagueness endemic to guid- ance on ethical disclosure and AI—what consti- tutes an appropriate manner of communication? There is a growing awareness of the importance of communicating AI system function and perfor- mance clearly and understandably. This is both a matter of public interest, for mitigating harms caused by, for example, racial biases in the perfor- mance of human classifiers (Raji and Buolamwini, 2019), but also in the interest of system providers, as users who feel like they do not understand how machine learning models work and what kind of information they rely on tend to be more resistant to using them (Poursabzi-Sangdeh et al., 2021). However, while fairness (Dwork et al., 2018; Harrison et al., 2020) and privacy as general prin- ciples (Ji et al., 2014; Papernot et al., 2016) have been well-studied in the context of AI, transparency does not receive as much attention. The term transparency is overloaded (Lipton and Steinhardt, 2019), with many different definitions in the lit- erature (Felzmann et al., 2019). It has a broad meaning in the public consciousness, from which some studies adopt a “you’ll know it when you see it” definition (Doshi-Velez and Kim, 2017). As op- posed to the notions of transparency as explainabil- ity (Veale et al., 2018) and as invisibility (Hamilton et al., 2014), transparency as disclosure (Suzor et al., 2019), the …

Content truncated to 500 words