Improving Information Extraction with Learned Queries
Better extraction came from better questions, not just bigger models.
The paper reports an 18.6-point F1 gain from improving query design across four clinical benchmarks and five LLMs. It introduces LoQ for generating document-specific question sets and FeedQ for refining those questions against extraction results. Fine-tuned 4B models using optimized questions matched or beat expert-derived baselines and outperformed much larger untuned models. The authors also released 12,820 optimized questions. ArXiv · AI/CL/LG's note
The paper reports an 18.6-point F1 gain from improving query design across four clinical benchmarks and five LLMs. It introduces LoQ for generating document-specific question sets and FeedQ for refining those questions against extraction results. Fine-tuned 4B models using optimized questions matched or beat expert-derived baselines and outperformed much larger untuned models. The authors also released 12,820 optimized questions. ArXiv · AI/CL/LG's note
score 5