RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL (2302.05965v3)

Abstract: One of the recent best attempts at Text-to-SQL is the pre-trained LLM. Due to the structural property of the SQL queries, the seq2seq model takes the responsibility of parsing both the schema items (i.e., tables and columns) and the skeleton (i.e., SQL keywords). Such coupled targets increase the difficulty of parsing the correct SQL queries especially when they involve many schema items and logic operators. This paper proposes a ranking-enhanced encoding and skeleton-aware decoding framework to decouple the schema linking and the skeleton parsing. Specifically, for a seq2seq encoder-decode model, its encoder is injected by the most relevant schema items instead of the whole unordered ones, which could alleviate the schema linking effort during SQL parsing, and its decoder first generates the skeleton and then the actual SQL query, which could implicitly constrain the SQL parsing. We evaluate our proposed framework on Spider and its three robustness variants: Spider-DK, Spider-Syn, and Spider-Realistic. The experimental results show that our framework delivers promising performance and robustness. Our code is available at https://github.com/RUCKBReasoning/RESDSQL.

Decoupling Schema Linking and Skeleton Parsing in Text-to-SQL with RESDSQL

The paper "RESDSQL: Decoupling Schema Linking and Skeleton Parsing for Text-to-SQL" addresses critical challenges in the task of translating natural language questions into SQL queries, a nuanced problem in the domain of NLP and database management. By decoupling two core components of the Text-to-SQL process—schema linking and skeleton parsing—this research puts forth a novel framework aimed at enhancing both the performance and robustness of these systems.

Motivation and Challenges

In Text-to-SQL models, particularly those relying on sequence-to-sequence (seq2seq) architectures, generating SQL queries from natural language is complicated by the need to integrate both database schema elements and SQL structures. This entanglement often increases parsing difficulty, especially in queries with numerous schema items and logical operators. This paper identifies these intertwined processes as a crucial bottleneck and suggests a groundbreaking separation, hypothesizing it could simplify query parsing and improve outcomes.

Proposed Framework

The authors introduce RESDSQL, a framework that decouples schema linking from skeleton parsing through a sequential process involving ranking-enhanced encoding and skeleton-aware decoding.

1. Ranking-Enhanced Encoding: This phase involves refining the input to the encoder by ranking and filtering schema items using a pre-trained cross-encoder. This cross-encoder classifies schema elements based on their relevance to the input natural language question, thereby streamlining the encoding process by incorporating only the most pertinent database schema components.
2. Skeleton-Aware Decoding: The decoder subsequently generates an intermediate SQL skeleton before forming the final SQL query. This two-step decoding effectively constrains the more complex query generation by first establishing a simpler framework of SQL operation order, thus facilitating the subsequent fill-in of schema details.

Methodology Details

The construction of RESDSQL leverages innovations in the encoding and decoding processes. For encoding, a cross-encoder identifies and injects relevant schema items, reducing schema linking complexity. The cross-encoder is trained with a focus on table and column relevance, employing focal loss to counter class imbalance and improve classification accuracy.

In decoding, SQL skeleton parsing is introduced to simplify the subsequent SQL generation. This approach benefits from a sequential generation pipeline, where parsing difficulties are reduced by initially focusing on the overarching structure (skeleton), and then iteratively filling in the content.

Results and Implications

The framework was evaluated on the Spider dataset, a challenging benchmark in Text-to-SQL translation, and recorded significant improvements in both Exact-match (EM) and EXecution (EX) metrics compared to existing models, including T5-based models enhanced with the PICARD grammar-based decoder. Also, RESDSQL demonstrated robust performance on Spider's variants designed to simulate realistic adversities, highlighting its robustness to schema modifications.

The proposed decoupling approach introduces a new paradigm to the Text-to-SQL task, suggesting that similar decoupling strategies could be beneficial in other semantic parsing tasks or frameworks encountering complex interdependencies. The simplification of parsing tasks implicitly achieved through such decoupling could further practical applications in AI by making natural language database querying more accessible to non-expert users.

Future Directions

The framework's promising results invite further exploration into expanding the decoupling strategy to other complex parsing tasks within NLP. Future work might investigate adaptive filtering mechanisms for schema item selection, optimizing skeleton generation techniques, or extending the framework to accommodate more complex SQL functionalities. Additionally, exploring the framework's adaptability to diverse datasets and domain-specific databases could further establish the utility of the decoupling approach.

In conclusion, RESDSQL advances the domain of Text-to-SQL translation by effectively decoupling schema linking and skeleton parsing, offering a significant leap in addressing the intricacies of natural language to SQL conversion through innovative encoding and decoding strategies.