How Salesforce’s Seq2SQL Turns Natural Language into SQL with Reinforcement Learning

Seq2SQL: Natural‑Language to SQL Generation

The paper Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning proposes a sequence‑to‑sequence model that translates a user’s natural‑language question into a SQL query. Because a single question can correspond to many syntactically different but semantically equivalent SQL statements, the authors augment the model with reinforcement learning. The reward is defined as the execution accuracy: the generated query receives a positive reward if it returns the same result as a reference query, regardless of exact syntax.

Model Architecture and Training

Encoder processes the tokenized natural‑language question.

Decoder generates the SQL statement token by token, conditioned on the encoder hidden states.

Column‑attention mechanism aligns question words with database schema column names, reducing the need to enumerate all possible column tokens.

Reinforcement learning (policy gradient) is applied after supervised pre‑training; the reward is computed by executing the generated query on the target database.

Vocabulary Constraints to Simplify Translation

To keep the learning problem tractable, the authors limit the vocabulary size in three places:

Words appearing in the user question.

Tokens representing database schema elements (table and column names).

Tokens that can appear in the target SQL query.

These constraints reduce the combinatorial explosion that occurs as questions become longer and schemas become larger.

Privacy Considerations

The authors note that models which directly access production databases could inadvertently expose sensitive information (e.g., queries involving personal identifiers). Deployments therefore need safeguards such as query sanitization, access control, or on‑device inference.

WikiSQL Dataset

To train and evaluate Seq2SQL, the researchers released the WikiSQL dataset, the largest publicly available NL‑SQL corpus at the time. The construction pipeline consists of:

Harvesting HTML tables from Wikipedia.

Generating random SQL queries for each table using a SQL generator that respects table schema.

Creating natural‑language questions by crowdsourcing on Amazon Mechanical Turk; each generated query is paraphrased by multiple workers.

Double‑checking each paraphrase by dedicated annotators to ensure quality.

The resulting dataset contains over 80,000 examples, each consisting of a table, a SQL query, and one or more natural‑language questions.

Illustrative Example

Given the question “Which team ranked first in the college football tour?” the system produces a SQL query such as:

SELECT team_name FROM rankings WHERE sport = 'football' ORDER BY rank ASC LIMIT 1;

Executing the query returns the answer (e.g., “University of Michigan”).