Summary: Zero-Shot Query Reformulation for Conversational Search
- Conversational search is gaining popularity with the rise of voice assistants, but suffers from lack of training data. Supervised methods require large labeled datasets which are expensive to obtain.
- Recent work has focused on zero-shot conversational search to avoid dependence on labeled data. However, existing methods have limitations: tied to specific models, lack explainability, and struggle with omitted information.
- This paper proposes a novel Zero-shot Query Reformulation (ZeQR) framework to address these limitations. It leverages machine reading comprehension models to resolve coreference and omission ambiguities in queries without conversational search supervision.
- ZeQR converts the query reformulation task into reading comprehension by posing template questions to identify ambiguous pronouns and fill in omitted information. This makes it model-agnostic and explains how it enhances query understanding.
- Experiments on TREC CAsT datasets show ZeQR consistently outperforms state-of-the-art zero-shot baselines. Analysis reveals the performance boost comes from resolving omission ambiguities.
- Key advantages of ZeQR are its versatility for any retriever, improved explainability, and effectively handling omissions. It shifts dependency from scarce conversational datasets to abundant reading comprehension data.
- Future work includes exploring different MRC datasets, reducing computational complexity, and testing with large language models like ChatGPT which achieved near human-level performance.
This paper introduces a novel zero-shot query reformulation approach for conversational search that leverages reading comprehension to resolve ambiguities without need for labeled conversational data. Key innovations are its model-agnostic nature, explainability, and omission handling.
disclosure: the Author uses AI to generate rough draft summaries.
