CoXQL: A Dataset for Parsing Explanation Requests in Conversational XAI Systems

Intent recognition is critical for explainable AI systems: The paper identifies that conversational XAI systems need to accurately understand what users are asking about AI decision-making, covering 31 different types of explanation requests—a challenge that existing systems struggle with due to limited training data in this specialized domain.

Current LLMs struggle with complex explanation requests: The research finds that large language models have particular difficulty when users ask for explanations involving multiple pieces of information simultaneously (multi-slot intents), suggesting that even advanced AI systems need improvement in parsing nuanced questions about how other AI systems work.

Structured parsing outperforms free-form conversation: The paper confirms that intent-based systems using template validation (their "MP+" approach) are more reliable than allowing users to ask questions in completely open-ended ways, indicating that some constraints on how users interact with XAI systems may actually improve their ability to get accurate explanations.

A new benchmark dataset addresses a gap in XAI research: CoXQL is presented as the first dataset specifically designed for training systems to understand explanation requests in conversational settings, providing a foundation for future development of more reliable explainable AI interfaces.

Explainability interfaces require specialized training: The findings suggest that general-purpose language models need domain-specific datasets and approaches to effectively help users understand AI decision-making, rather than relying solely on their general capabilities.