Beyond Isolated Dots: Benchmarking Structured Table Construction as Deep Knowledge Extraction