Interactive question learning improves person search with partial clues
What Should We Ask Next? Retrieval-Aware Question Learning under Partial Evidence
Artificial Intelligence
Summary
Finding a person in a large group using just partial clues is hard. The authors worked on teaching a system how to pick the best questions to ask that will give the most helpful clues next. Their method, called RAVEL, learns from feedback during the search to ask better questions over time. This makes the system better at identifying people even when clues are missing or unclear. The system especially improves on tough searches by focusing on detailed, open-ended questions.
What this means in practice
- •For security system developers: Create person search tools that ask adaptive questions to improve identification under incomplete information.$Commercial implications: Enables development of advanced security and surveillance products with interactive search capabilities for faster and more accurate person detection.
- •For customer support teams: Design chatbots that learn to ask better follow-up questions to quickly resolve customer issues when initial information is limited.
Authors
Lyucheng Qian, John Yuehan Zhang, Pingyu Wang
Abstract
Interactive retrieval under partial evidence is a sequential information-acquisition problem: an agent must decide which question will create the most useful evidence for the next retrieval update. Existing systems train this decision by imitating an offline ordering of candidate QA pairs, although question value is determined by the response it elicits and its downstream effect on retrieval. We establish that candidate discriminativeness and perceived usefulness provide weak supervision for this objective, then introduce RAVEL, a retrieval-aware online reinforcement learning framework for interactive person re-identification. RAVEL initializes from supervised question generation, observes the current Top-4 candidates directly, and optimizes the question policy with rank feedback from the full question-answer-retrieval loop. Experiments on Interactive-PEDES show that RAVEL delivers progressively stronger retrieval performance across five interaction rounds. Further analysis shows that RAVEL reallocates the questioning budget toward localized open-ended attributes, which provide more useful retrieval evidence and yield the largest gains on initially difficult queries.