What Should We Ask Next? Retrieval-Aware Question Learning under Partial Evidence
RAVEL trains the next-question policy on retrieval feedback, not an offline guess at which question should come first.
The paper frames interactive retrieval as a loop where a question matters only if its answer improves the next ranking. RAVEL starts from supervised question generation, looks at the current Top-4 candidates, and updates its policy through rank feedback from the full question-answer-retrieval process. On Interactive-PEDES, the authors report stronger retrieval performance over five interaction rounds. The gains are tied to asking more localized, open-ended attribute questions, especially for initially difficult queries. ArXiv · AI/CL/LG's note
The paper frames interactive retrieval as a loop where a question matters only if its answer improves the next ranking. RAVEL starts from supervised question generation, looks at the current Top-4 candidates, and updates its policy through rank feedback from the full question-answer-retrieval process. On Interactive-PEDES, the authors report stronger retrieval performance over five interaction rounds. The gains are tied to asking more localized, open-ended attribute questions, especially for initially difficult queries. ArXiv · AI/CL/LG's note
score 5