ConvXAI: Delivering Heterogeneous AI Explanations via Conversations to Support Human-AI Scientific Writing (2305.09770v6)

Abstract: Despite a surge collection of XAI methods, users still struggle to obtain required AI explanations. Previous research suggests chatbots as dynamic solutions, but the effective design of conversational XAI agents for practical human needs remains under-explored. This paper focuses on Conversational XAI for AI-assisted scientific writing tasks. Drawing from human linguistic theories and formative studies, we identify four design rationales: "multifaceted", "controllability", "mix-initiative", "context-aware drill-down". We incorporate them into an interactive prototype, ConvXAI, which facilitates heterogeneous AI explanations for scientific writing through dialogue. In two studies with 21 users, ConvXAI outperforms a GUI-based baseline on improving human-perceived understanding and writing improvement. The paper further discusses the practical human usage patterns in interacting with ConvXAI for scientific co-writing.

Delivering Heterogeneous AI Explanations via Conversations for Human-AI Scientific Writing

The paper explores the design and application of conversational eXplainable AI (XAI) systems to enhance human-AI collaboration, particularly in the field of scientific writing. The authors present a paper focused on designing conversational XAI interfaces that can dynamically cater to human needs by providing context-aware, multifaceted explanations.

Key Aspects of the Research

The authors identify four essential design principles for effective conversational XAI systems:

1. Multifaceted Explanation: The system should provide explanations covering diverse aspects that users might inquire about. This involves providing various types of explanations, such as feature attribution, model confidence, and similar examples, tailored to the needs of different users.
2. Controllability: Users should be able to control the level of detail and type of explanations they receive. This involves allowing users to customize how much information they want and which specific aspects of the model's behavior they wish to explore.
3. Mixed-Initiative Interaction: The system should allow both the user and the AI to drive the conversation. This might include the AI proactively suggesting possible next steps or additional explanations.
4. Context-Aware Drill-Down: Users should be able to dig deeper into explanations in a context-aware manner. The system should track the interaction history to provide coherent and relevant follow-up information.

Methodology and Implementation

A prototype system, referred to as ConvXAI, was developed to embody these design principles. This system incorporates two AI models for assisting writing: one focusing on writing structure and the other on writing style. The system provides a dialogue interface where users can iteratively refine their scientific abstracts with the help of AI-generated feedback and explanations.

The methodology involved formative studies to gather user needs, followed by a prototype development phase informed by those needs. The system's effectiveness was evaluated in two user studies involving scientific writing tasks.

Findings and Implications

The results from user studies indicate that the conversational approach in ConvXAI helps users better understand and utilize AI feedback to improve their scientific writing. The findings highlight the advantages of a dynamic conversational interface over traditional GUI-based systems, particularly in terms of user engagement and satisfaction.

The implication of this research is significant for the development of future AI systems that need to provide transparent, understandable, and user-centric explanations. The paper suggests that incorporating conversational design elements into XAI systems can significantly enhance user experience and satisfaction, especially in complex task domains such as scientific writing.

Future Directions

The research opens several avenues for future exploration. One potential direction is to extend the application of conversational XAI systems beyond scientific writing to other domains where user understanding of AI decisions is critical. Additionally, there is room for further exploration into integrating more nuanced customizations and broader XAI methods into the conversational framework.

Overall, the paper provides a structured approach to designing conversational XAI systems that effectively bridge the gap between AI capabilities and human interpretative needs, paving the way for more effective human-AI collaboration.