Interactive Model Cards: A Human-Centered Approach to Model Documentation (2205.02894v1)

Abstract: Deep learning models for NLP are increasingly adopted and deployed by analysts without formal training in NLP or ML. However, the documentation intended to convey the model's details and appropriate use is tailored primarily to individuals with ML or NLP expertise. To address this gap, we conduct a design inquiry into interactive model cards, which augment traditionally static model cards with affordances for exploring model documentation and interacting with the models themselves. Our investigation consists of an initial conceptual study with experts in ML, NLP, and AI Ethics, followed by a separate evaluative study with non-expert analysts who use ML models in their work. Using a semi-structured interview format coupled with a think-aloud protocol, we collected feedback from a total of 30 participants who engaged with different versions of standard and interactive model cards. Through a thematic analysis of the collected data, we identified several conceptual dimensions that summarize the strengths and limitations of standard and interactive model cards, including: stakeholders; design; guidance; understandability & interpretability; sensemaking & skepticism; and trust & safety. Our findings demonstrate the importance of carefully considered design and interactivity for orienting and supporting non-expert analysts using deep learning models, along with a need for consideration of broader sociotechnical contexts and organizational dynamics. We have also identified design elements, such as language, visual cues, and warnings, among others, that support interactivity and make non-interactive content accessible. We summarize our findings as design guidelines and discuss their implications for a human-centered approach towards AI/ML documentation.

A Professional Exploration of Interactive Model Cards for Model Documentation

The paper "Interactive Model Cards: A Human-Centered Approach to Model Documentation" addresses the gap between the technical creation and utilization of machine learning models, particularly in NLP, and the understanding and accessibility for non-expert users. The increasing complexity of deep learning models requires sophisticated documentation to ensure these tools are used appropriately and effectively by individuals without a formal background in machine learning or NLP. This research introduces interactive model cards (IMCs) as an evolution of the traditional static model cards designed to address this gap.

Interactive model cards aim to provide users with dynamic components to explore model documentation and engage with models interactively. The paper applies a human-centered design approach, incorporating insights from both experts and non-expert analysts. The research is conducted in two phases: an initial conceptual paper using design probes and interviews with experts, followed by an evaluative paper with non-expert analysts using a functional IMC prototype.

Conceptual Foundations and Design Guidelines

Initially, the authors engaged with experts in machine learning, ethics, and NLP to delineate the conceptual scope and design guidelines for IMCs. The experts critiqued static model cards for their failure to cater to non-experts, and they suggested that interactive components could be beneficial for data exploration and conceptual understanding. Crucially, the experts emphasized the importance of thoughtful information hierarchy, language clarity, visual design, and interactivity. These factors ensure that model cards can communicate effectively.

Design guidelines emerging from this phase include careful structuring of information, leveraging interaction to foster users' understanding, providing actionable guidance, and incorporating defaults to promote skepticism and critical thinking. The thematic exploration identified stakeholders’ needs, and interpreted that stakeholders require language simplification, improved information design, and better ways to contextualize data. These design insights were incorporated into the next development phase.

Evaluation with Non-Expert Analysts

Building upon the initial conceptual design, a functional prototype for an IMC was developed and assessed by non-expert analysts. This prototype incorporated interactive elements such as data visualization, options to upload data, and explore sensitive population sub-groups. Participants noted that the IMC provided a user's overview of model behavior in contrast to static model cards, which were often verbose and technical. The interactive component was favored for allowing experimentation, which surfaced insights into model capabilities and limitations that were not immediately apparent in static documentation.

Interactive components increased the utility of model cards by bridging the gap between sophisticated model capabilities and user understanding, enhancing the potential for productive skepticism. Analysts appreciated the ability to test and contest the model's outputs, indicating this interaction reassured them about model capabilities and limitations. The participants highlighted the IMC's facilitation of sharing information across organizational spectrums, suggesting that the enhanced clarity and interactivity improved communication with various stakeholders.

Implications and Future Directions

The research exemplifies how IMCs can scaffold an improved human-centered approach in AI/ML documentation. By allowing non-experts to interrogate and understand model behavior more comprehensively, IMCs can play a critical role in ensuring ethical deployment and use of machine learning models in diverse contexts. Furthermore, the insights gathered suggest the need for integrating such documentation formats into existing organizational processes, potentially modifying how AI/ML models are adopted and regulated.

This paper opens avenues for further research into optimizing IMC design to balance detail with simplicity and integrating it with educational resources to enhance user understanding of complex machine learning systems. Additionally, the research underscores the challenges in surfacing ethical considerations in the documentation in ways that are both informative and practically actionable.

In conclusion, the paper makes a compelling case for interactive model cards as a means to meet the needs of a broader range of stakeholders in the AI ecosystem. The thoughtful blend of interactivity and static information can enhance the usability of model documentation for non-experts, ensuring these powerful tools are harnessed effectively and ethically.