Track-SQL: Enhancing Generative Language Models with Dual-Extractive Modules for Schema and Context Tracking in Multi-turn Text-to-SQL
arXiv:2603.05996v1 Announce Type: new Abstract: Generative language models have shown significant potential in single-turn Text-to-SQL. However, their performance does not extend equivalently to multi-turn Text-to-SQL. This is primarily due to generative language models' inadequacy in handling the complexities of context information and dynamic schema linking in multi-turn interactions. In this paper, we propose a framework named Track-SQL, which enhances generative language models with dual-extractive modules designed to track schema and contextual changes in multi-turn Text-to-SQL. Specifically, Track-SQL incorporates a \emph{Semantic-enhanced Schema Extractor} and a \emph{Schema-aware Context Extractor}. Experimental results demonstrate that Track-SQL achieves state-of-the-art performance on the SparC and CoSQL datasets. Furthermore, detailed ablation studies reveal that Track-SQL significantly improves execution accuracy in multi-turn interactions by 7.1\% and 9.55\% on these data
arXiv:2603.05996v1 Announce Type: new Abstract: Generative language models have shown significant potential in single-turn Text-to-SQL. However, their performance does not extend equivalently to multi-turn Text-to-SQL. This is primarily due to generative language models' inadequacy in handling the complexities of context information and dynamic schema linking in multi-turn interactions. In this paper, we propose a framework named Track-SQL, which enhances generative language models with dual-extractive modules designed to track schema and contextual changes in multi-turn Text-to-SQL. Specifically, Track-SQL incorporates a \emph{Semantic-enhanced Schema Extractor} and a \emph{Schema-aware Context Extractor}. Experimental results demonstrate that Track-SQL achieves state-of-the-art performance on the SparC and CoSQL datasets. Furthermore, detailed ablation studies reveal that Track-SQL significantly improves execution accuracy in multi-turn interactions by 7.1\% and 9.55\% on these datasets, respectively. Our implementation will be open-sourced at https://github.com/DMIRLAB-Group/Track-SQL.
Executive Summary
The article introduces Track-SQL, a novel framework that addresses the limitations of generative language models in multi-turn Text-to-SQL by introducing dual-extractive modules—specifically a Semantic-enhanced Schema Extractor and a Schema-aware Context Extractor. These modules are designed to improve context and schema tracking across conversational turns. The experimental evaluation on SparC and CoSQL datasets demonstrates significant gains, with 7.1% and 9.55% improvements in execution accuracy, respectively, establishing Track-SQL as a state-of-the-art solution. The work fills a critical gap in adapting generative models for dynamic, multi-turn SQL generation environments.
Key Points
- ▸ Introduction of dual-extractive modules to enhance context and schema tracking
- ▸ Demonstrated performance improvements on SparC and CoSQL datasets
- ▸ Open-source availability of the implementation
Merits
Innovation
Track-SQL introduces a targeted architectural solution to a specific deficiency in generative models for multi-turn SQL, offering measurable gains in accuracy.
Empirical Validation
The framework’s effectiveness is substantiated through rigorous experimental validation on benchmark datasets, lending credibility to the claims.
Demerits
Scope Constraint
The study focuses on specific benchmark datasets; broader applicability to diverse SQL domains or user interfaces remains unverified.
Complexity
Integrating dual-extractive modules may increase computational overhead or model complexity, potentially affecting scalability or real-time deployment.
Expert Commentary
Track-SQL represents a meaningful advancement in the intersection of generative AI and database query systems. The dual-extractive architecture is a sophisticated response to a well-documented limitation: the inability of vector-based models to maintain coherent context and schema alignment across multi-turn conversations. The specificity of the modules—Semantic-enhanced Schema Extractor and Schema-aware Context Extractor—suggests a deep understanding of the underlying challenges in mapping conversational intent to SQL syntax. Moreover, the measurable improvements in execution accuracy (7.1% and 9.55%) are statistically significant and reflect a substantive impact. While the open-source availability promotes reproducibility and adoption, the long-term success will depend on scalability and applicability beyond the tested datasets. Notably, the paper avoids overgeneralization and remains grounded in empirical evidence, which strengthens its scholarly credibility. This work sets a new benchmark for evaluating performance in multi-turn SQL generation and should influence future research directions in conversational database interfaces.
Recommendations
- ✓ Researchers should extend Track-SQL’s architecture to other query domains beyond SQL, such as NoSQL or graph databases.
- ✓ Industry stakeholders should evaluate integrating Track-SQL into production-grade conversational AI platforms, particularly those requiring robust context and schema awareness.