BROWSE
Detail
Conference Paper
Cloze Your Gaps: Improving Knowledge Extraction via Query Rewriting
Cited 0 time in 
Share 
Share
- Authors
- Sanghyeon Cho, Myeongjin Lee, Jeong Heo, Jong-hun Shin, Kiyoung Lee, Soojong Lim
- Issue Date
- 2026-06
- Citation
- Pacific-Asia Conference on Knowledge Discovery and Data Mining (PAKDD) 2026, pp.204-215
- Publisher
- Springer
- Language
- English
- Type
- Conference Paper
- Abstract
- Open-domain question answering with large language models typically adopts a retrieve-then-read pipeline, which is limited by the retriever’s accuracy and its inability to fully exploit the model’s parametric knowledge. Recent work has explored generate-then-read approaches, where the LLM generates supporting contexts directly. In this paper, we propose a novel pipeline that enhances generate-then-read by introducing a query rewriting module that transforms open-domain questions into cloze-style prompts. These prompts better align with the pretraining distribution of base models and improve the quality of generated contexts. To address variability in model responses to different cloze formulations, we further introduce a query sampling procedure that generates multiple cloze variants per question. The resulting contexts are aggregated using three strategies—SELECT, FUSE, and CONCAT—and filtered via lightweight retrievers, rerankers, and LLM-based ranking prompts. Extensive experiments demonstrate that cloze-based query rewriting enhances knowledge extraction in both base and continuously pretrained models. This work highlights the efficacy of leveraging cloze-style prompts and context aggregation for robust and adaptable open-domain QA systems.
- Keyword
- Query Rewriting, Generate-then-read, Knowledge Extraction
- KSP Keywords
- Knowledge extraction, Language Models, Query Rewriting, Query Sampling, Question Answering, Sampling procedure