Bridge-inventory question answering benchmark: 300 Korean queries with ground-truth answers (2011-2024)
收藏资源简介:
Benchmark used to evaluate natural-language analytics over Korea's national bridge inventory. The inventory consists of 14 annual snapshot files (2011-2024) totalling 474,670 bridge-year records; the 2024 snapshot (40,006 records) is the default reference year. questions_300.json contains 300 queries with category, complexity level, the Korean original, and an English reference translation. All queries were issued in Korean; the translations are for reference only. ground_truth_300.json contains the expected answer for each query together with the interpreted intent slots. Every expected value was recomputed deterministically from the published annual files. The two files join on "id". Queries cover aggregation, filtering, ranking, comparison, ratio, specific-facility lookup, data quality, and multi-type requests. The annual inventory files are published by the Korea Institute of Civil Engineering and Building Technology (KICT); this deposit contains only the derived benchmark and ground truth.



