Transparent and Interpretable Conversational Agents

Smart-speakers including Amazon Alexa and Google Home are emerging as a key consumer technology. Approximately 16% of adults (39 million) in the United States now own one of these devices. As such, these devices can shape the future of user interactions and engagements. However, current devices have a range of usability issues, which negatively impact users’ experiences and satisfactions. Specifically, interactions with these devices often fail to: i) support discoverability and learnability — users are unable to explore and understand the capabilities of these devices, which limits their usefulness; ii) provide adequate explanation regarding what caused an error and how a user can rectify it — a key usability issue which can lead to user frustrations and disengagement. These issues stem from the lack of interpretability of underlying system, which then can result in incorrect mental models for users.

In this project, we are aiming to address these issues by improving the interpretability of smart-speakers and thus, leading to better user mental models. Specifically, we will adopt a user-centric approach to interpretability by determining both content (“what to explain”) and presentation (“how to explain”) to users. While there has been recent work on interpretability of complex systems, smart-speakers have unique constraints and challenges (e.g., prior work on using visualization for explaining decisions will not work for voice-based interactions). Addressing these challenges can improve quality of user interactions as well as extend the capabilities of these systems to support complex user tasks.