A View-Oriented Skeleton Generation Method for Improving Multi-Table Text-to-SQL Translation Accuracy

Recently, research on natural language to structured query (Text-to-SQL) has advanced rapidly, aiming to enable models to automatically translate natural language queries into executable SQL statements. When handling certain categories of multi-table queries, specifically those that can be abstracted as single‑table or natural‑join views, models must not only generate SQL syntax but also correctly infer join relationships. To substantially reduce generation difficulty and improve translation accuracy for these query types, this paper proposes a view‑level semantic abstraction‑based SQL generation method (View‑SQL). It first generates an SQL skeleton on a pre-built view, after which a rule-based module completes the table names and the FROM clause to produce the final executable SQL statement. To achieve this, View-SQL introduces two collaborative processing paths. The first is the View Path (Text2SQLSkeleton), where a View Classifier determines which type of view the natural language query corresponds to, and a T5-based decoder generates the SQL skeleton for that view. The second is the Non-View Path (Text2SQL), which is invoked when a query cannot be mapped to any existing view. In this path, the model adopts BERT and multi-head attention mechanisms to select relevant tables and columns before generating the SQL statement. Through this hierarchical framework, View-SQL effectively reduces the complexity of multi-table inference. Experimental results demonstrate that on our validation split, since the Spider test set is not publicly available, View-SQL achieves accuracy improvements of 1.55% and 2.13% on the Spider-syn and Spider EM datasets, respectively.