beader/tianchi_nl2sql

Implements a two-stage BERT-based architecture that decomposes SQL generation into distinct prediction tasks: Model 1 predicts selected columns and WHERE clause conditions using Chinese BERT-wwm embeddings with column type tokens, while Model 2 enumerates candidate condition-value pairs and frames them as binary classification problems. The approach reformulates the SQL output structure by merging aggregation operations with column selection and separating condition operators from values, enabling more granular modeling of each SQL component.

559 stars. No commits in the last 6 months.