Conceitos Básicos
SQLformer is a novel Transformer-based architecture that generates SQL queries as abstract syntax trees in an autoregressive manner, incorporating structural inductive bias to improve performance on text-to-SQL translation tasks.
Resumo
The paper introduces SQLformer, a novel Transformer-based model for text-to-SQL translation. The key highlights are:
SQLformer incorporates learnable table and column token embeddings in the encoder to select the most relevant database schema elements for a given natural language question. This schema-aware question representation is then used as input to the decoder.
The SQLformer decoder extends the original Transformer decoder by integrating node type, adjacency, and previous action embeddings to generate SQL queries autoregressively as a sequence of actions derived from a SQL grammar.
Comprehensive experiments show that SQLformer achieves state-of-the-art performance on five widely used text-to-SQL benchmarks, including Spider, SParC, and CoSQL. It particularly excels on complex queries and demonstrates strong zero-shot domain generalization capabilities.
Ablation studies confirm the importance of the table and column selection mechanism, as well as the benefits of the Transformer-based decoder compared to previous LSTM-based approaches.
Overall, the paper presents a novel and effective Transformer-based architecture for text-to-SQL translation, addressing key challenges such as domain generalization and complex query generation.
Estatísticas
The paper does not provide any specific numerical data or statistics. The key results are reported in terms of Exact Match (EM) accuracy on various text-to-SQL benchmarks.
Citações
"SQLformer, a novel Transformer architecture specifically crafted to perform text-to-SQL translation tasks."
"Our model predicts SQL queries as abstract syntax trees (ASTs) in an autoregressive way, incorporating structural inductive bias in the encoder and decoder layers."
"Comprehensive experiments show the state-of-the-art performance of SQLformer across five widely used text-to-SQL benchmarks."