Seq2SQL: Generating Structured Queries from Natural Language using Reinforcement Learning

A significant amount of the world’s knowledge is stored in relational databases. However, the ability for users to retrieve facts from a database is limited due to a lack of understanding of query languages such as SQL. We propose Seq2SQL, a deep neural network for translating natural language questions to corresponding SQL queries. Our model leverages the structure of SQL queries to significantly reduce the output space of generated queries. Moreover, we use rewards from in-the-loop query execution over the database to learn a policy to generate unordered parts of the query, which we show are less suitable for optimization via cross entropy loss. In addition, we will publish WikiSQL, a dataset of 80654 hand-annotated examples of questions and SQL queries distributed across 24241 tables from Wikipedia. This dataset is required to train our model and is an order of magnitude larger than comparable datasets. By applying policy-based reinforcement learning with a query execution environment to WikiSQL, our model Seq2SQL outperforms attentional sequence to sequence models, improving execution accuracy from 35.9% to 59.4% and logical form accuracy from 23.4% to 48.3%.

💡 Research Summary

The paper introduces Seq2SQL, a neural architecture that translates a natural‑language question together with a table schema into a correct SQL query. The authors observe two major shortcomings of conventional sequence‑to‑sequence (Seq2Seq) semantic parsers for this task: (1) the output vocabulary is unnecessarily large because SQL queries are composed almost entirely of a small set of keywords, column names, and words from the question; (2) the WHERE clause often contains multiple predicates whose order does not affect the query result, making cross‑entropy supervision inappropriate.

To address (1), Seq2SQL builds an augmented input sequence that concatenates column names, a limited SQL keyword vocabulary, and the question tokens, each delimited by sentinel tokens. A bidirectional LSTM encodes this sequence, and a pointer network selects tokens directly from the input, effectively restricting the decoder’s vocabulary to the union of columns, keywords, and question words. This dramatically shrinks the search space compared with a vanilla softmax decoder.

For (2), the model decomposes a SQL query into three logical components that mirror the grammar of SELECT statements: (i) an aggregation operator (COUNT, MIN, MAX, or NULL), (ii) the SELECT column, and (iii) the WHERE clause. The aggregation and SELECT components are trained with standard cross‑entropy loss because they have a single correct token. The WHERE clause, however, is generated by a pointer‑based decoder that samples tokens step‑by‑step. After a full WHERE clause is produced, the complete SQL query is executed against the underlying database. The execution result yields a scalar reward: +1 for a correct result, –1 for an incorrect result, and –2 for an invalid SQL syntax. Using the REINFORCE policy‑gradient algorithm, the model maximizes the expected reward, thereby learning to produce unordered predicate sets without penalizing permutations that are semantically equivalent.

The overall loss is a simple sum of the three sub‑losses (aggregation, selection, and WHERE), allowing simultaneous optimization of supervised and reinforcement signals.

A major contribution of the work is the release of WikiSQL, a large‑scale benchmark consisting of 80 654 (question, schema, SQL) triples derived from 24 241 Wikipedia tables. The dataset was created by automatically generating template questions, having crowd‑workers paraphrase them, and then verifying the paraphrases for semantic fidelity. WikiSQL is an order of magnitude larger than previous semantic‑parsing datasets such as GeoQuery or ATIS, and it includes realistic table structures, varying column counts, and diverse question types. The authors also provide the raw tables in JSON, a ready‑to‑use SQL database, and a query‑execution engine for research reproducibility.

Experimental results show that Seq2SQL outperforms the strongest prior model (Dong & Lapata’s attentional Seq2Seq) by a large margin: execution accuracy rises from 35.9 % to 59.4 %, and logical‑form (exact‑match) accuracy improves from 23.4 % to 48.3 %. An ablation that removes the reinforcement‑learning component (i.e., using only the augmented pointer network) achieves 53.3 % execution accuracy, confirming that the policy‑gradient training of the WHERE clause contributes a substantial gain.

The paper discusses limitations: the current system only handles simple binary predicates in the WHERE clause and does not support joins, sub‑queries, or more complex logical operators (OR, NOT). The reward function is coarse‑grained, providing only binary feedback about correctness, which may limit learning of nuanced query variations. Future work is suggested to extend the grammar coverage, incorporate multi‑table schemas, and design richer reward signals that capture partial correctness.

In summary, Seq2SQL demonstrates that explicitly modeling SQL’s hierarchical structure and employing reinforcement learning for unordered components can dramatically improve natural‑language‑to‑SQL translation. Coupled with the publicly released WikiSQL dataset, this work establishes a new benchmark and a solid foundation for subsequent research on neural semantic parsing for relational databases.