Learning to Clarify Underspecified Intents Under Limited Interaction

TL;DR

This paper introduces a reinforcement learning framework to help AI assistants clarify underspecified user intents by asking informative questions. The framework is tested in an image generation task and shows significant improvements in user utility recovery. It enables assistants to better disambiguate user intent with fewer questions and less interaction time.

Why It Matters

The proposed framework has practical significance as it can be applied to various AI assistant tasks, improving user experience and reducing interaction time. It can also be used to train language models to better understand user intent, leading to more effective human-computer interactions. This can have a significant impact on the development of more efficient and user-friendly AI systems.

Key Findings

• significant improvement in user utility recovery
• reduction in number of questions asked
• decrease in total interaction time