SQLNet

SQLNet is a neural network system designed to translate natural language questions into SQL queries that can be executed against relational databases. Introduced in 2019 in the text-to-SQL literature, SQLNet sought to improve cross-domain applicability and output validity by embedding a structured decoding process that aligns natural language with the database schema.

Architecture: SQLNet uses a skeleton-based, constrained decoding approach. It partitions SQL generation into components such as SELECT, FROM, WHERE, GROUP BY, and ORDER BY. For each component, a dedicated encoder encodes the question and the database schema, and a slot-filling decoder selects elements (columns, tables, operators, and literal values). Attention mechanisms guide choices by focusing on relevant parts of the question and schema, while constraints ensure syntactic and semantic validity of the resulting query.

Training and evaluation: The model was evaluated on cross-domain benchmarks such as the Spider dataset. Training uses pairs of natural language questions and corresponding SQL queries, with schema information as input. Performance is reported in exact-match accuracy and component-wise metrics, often showing gains over earlier sequence-to-SQL models by enforcing structural constraints during decoding.

Limitations and impact: SQLNet depends on explicit schema information and may struggle with complex, nested, or highly specialized queries. Generalization remains a challenge when encountering unseen schemas or non-standard SQL. Nevertheless, SQLNet influenced subsequent work on text-to-SQL by popularizing skeleton-based decoding and schema-aware representations, informing later models such as IRNet and related approaches.

See also: Text-to-SQL; Natural language interfaces to databases.