Semantic Parsing for Text-to-SQL: Datasets, Models, Evaluation, and Applications

Semantic parsing (Semantic Parsing) is a core natural‑language‑processing task that maps language to executable database queries; the Text‑to‑SQL sub‑task automatically converts a user’s natural‑language question into a SQL statement that can be run on a relational database.

The task consists of a parser, which receives a database schema and a question and outputs the corresponding SQL query, and an executor that runs the query; this article focuses on the parser, describing the mapping from question to database elements and the generation of syntactically correct SQL.

We summarize the evolution of Text‑to‑SQL datasets: early single‑domain corpora, multi‑domain collections such as Spider, CoSQL, NL2SQL, CSpider, and the newly constructed DuSQL dataset, which classifies data by domain count, single/multi‑table, question complexity, round‑trip count, and conversational extensions.

Model progress is organized into four categories: (1) Pointer‑Network‑based models (Seq2SQL, STAMP, Coarse2Fine, IRNet) that copy tokens from the input; (2) Sequence‑to‑set approaches (SQLNet, TypeSQL, SQLova, X‑SQL) that predict column‑keyword assignments; (3) Grammar‑based methods (TRANX) that generate abstract‑syntax‑tree actions; and (4) other enhancements such as graph‑neural‑network encoders (Global GNN, RATSQL) and execution‑guided decoding. Example generated SQL: Select 名称,所属省 From 中国城市 Where 绿化率 > 30% .

Evaluation uses two metrics: Exact Match (Acc qm ) which checks set‑wise equality of SQL components, and Execution Accuracy (Acc ex ) which measures whether the query returns the correct answer.

DuSQL was built by harvesting 813 tables from encyclopedic sources, clustering them into 200 databases, and semi‑automatically generating about 24 000 pairs using a grammar‑driven SQL generator followed by crowdsourced natural‑language paraphrasing.

DuParser combines three modules: component mapping (identifying table/column mentions), label identification (SQL‑keyword tagging via a sequence‑to‑set model), and grammar combination (a CYK‑style parser that assembles a syntax tree from a set of grammar rules, selecting the best rule by similarity between SQL fragments and question fragments). The framework offers interpretability, controllability, and domain‑agnostic generalization.

Real‑world deployments at Baidu include B2B customer‑service platforms (e.g., UNIT) and search‑engine scenarios, where Text‑to‑SQL powers structured Q&A, conversational query refinement, and enterprise table‑based search.

Remaining challenges are table‑recognition and normalization, incorporation of external knowledge for operations such as sorting or unit conversion, and extending the system to multi‑turn conversational settings that require context‑aware understanding.