bertrand-dr

Bertrand-DR (Bertrand Discriminative Re-ranker) is a 2020 research paper on improving neural text-to-SQL parsers for the Spider benchmark. Rather than generating SQL directly, the method trains a schema-agnostic BERT classifier to score and re-rank the candidate parses in a base model's beam output - recovering the cases where the correct query is in the beam but not at the top - and analyzes how the generator and re-ranker complement each other across query-hardness levels.

Co-authored at Got It AI R&D, the paper reached a top-4 score on the Spider leaderboard at the time of publication and is published on arXiv. Supporting preprocessing and log-parsing tooling lives in text2sql-utils.