RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL

One of the recent best attempts at Text-to-SQL is pre-trained language model. Due to structural property SQL queries, seq2seq model takes responsibility parsing both schema items (i.e., tables and columns) skeleton keywords). Such coupled targets increase difficulty correct queries especially when they involve many logic operators. This paper proposes a ranking-enhanced encoding skeleton-aware decoding framework decouple linking parsing. Specifically, for encoder-decode model, its encoder injected by most relevant instead whole unordered ones, which could alleviate effort during parsing, decoder first generates then actual query, implicitly constrain We evaluate our proposed on Spider three robustness variants: Spider-DK, Spider-Syn, Spider-Realistic. The experimental results show that delivers promising performance robustness. Our code available https://github.com/RUCKBReasoning/RESDSQL.