Schema linking is widely recognized as a key factor in improving text-to-SQL performance.Supervised fine-tuning approaches enhance SQL generation quality by explicitly finetuning schema linking as an extraction task.However, they suffer from two major limitations: (i) The training corpus of small language models restricts their cross-domain generalization ability.(ii) The extraction-based finetuning process struggles to capture complex linking patterns.To address these issues, we propose GenLink, a generation-driven schemalinking framework based on multi-model learning.Instead of explicitly extracting schema elements, GenLink enhances linking through a generation-based learning process, effectively capturing implicit schema relationships.By integrating multiple small language models, GenLink improves schema-linking recall rate and ensures robust cross-domain adaptability.Experimental results on the BIRD and Spider benchmarks validate the effectiveness of GenLink, achieving execution accuracies of 67.34% (BIRD), 89.7% (Spider development set), and 87.8% (Spider test set), demonstrating its superiority in handling diverse and complex database schemas.

Cross-domain Text-to-SQL necessitates the capability of semantic parsers to generalize to unseen databases, thus simplifying the process of creating natural language interfaces for databases. The existing Text-to-SQL parser exhibits limitations in its adaptability to new databases, and its execution accuracy is not sufficient for building conversational applications, typically necessitating further fine-tuning for specific databases. In this paper, we introduce Chat2DB, a conversational system designed for database interactions that enhances parser capabilities, rendering them applicable in real-world contexts. Within Chat2DB, we implement an interactive schema-ranking agent that optimizes the performance of LMs-based parsers cost-effectively. We further propose an adaptive retraining stage to allow trained Text-to-SQL parsers to quickly adapt to the target database. Experimental evaluations were conducted to validate the performance of the key components of Chat2DB. In the demonstration, we showcase the interactive visualization interface of Chat2DB to achieve more accurate querying of databases by natural language.

The goal of the Cross-domain Text-to-SQL task is to accurately translate natural language questions into executable SQL queries, even when applied to previously unseen databases. Recently, full fine-tuning of pretrained T5 models has achieved remarkable results. However, as the model size increases, this paradigm incurs significant computational costs. In this paper, we explore the effectiveness of existing parameter-efficient fine-tuning (PEFT) methods for training the T5 model on the Text-to-SQL task. Recognizing the limitations of PEFT methods in structure learning, we propose two structure-aware variants for Adapter and Prefix tuning methods, integrating a relational graph neural network within their architectures. We conducted extensive experiments to demonstrate the effectiveness of our proposed structure-aware adapter and structure-aware prefix tuning methods, which achieve performance comparable to the state-of-the-art T5-3B model while requiring only about 5.43% of the training parameters needed.

Bingfeng Chen, Shaobin Shi, Yongqi Luo, Boyan Xu, Ruichu Cai, Zhifeng Hao. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.