Vietnamese Text-to-SQL with Large Language Models: A Comprehensive Approach

Phu Nguyen Dac Hoang, Cuong Le Phuoc, Tri Doan Tran Cao, Nguyễn Minh Quang

Keywords: Text-to-SQL, Large Language Models, Few-shot, Chain-of-thought, Mini Schema, ViText2SQL, SQL

TL;DR: We use Vietnamese LLMs for text-to-SQL tasks, introducing few-shot learning, a chain-of-thought technique, and schema streamlining. Our method outperforms the state-of-the-art by 23% on ViText2SQL with one training epoch.

Abstract: In the current era of Artificial Intelligence (AI), the realm of database querying is experiencing a profound evolution. With the recent emergence of Large Language Models (LLMs), with a particular emphasis on Vietnamese in this study, a promising opportunity arises to bridge the gap between human language and database interactions. In this paper, we embark on realizing this vision through a three-pronged approach. Firstly, we introduce a few-shot learning method designed to enhance the database schema comprehension of Vietnamese LLMs. Secondly, we employ a chain-of-thought technique to systematically guide LLMs in capturing complex natural language expressions for SQL generation. Thirdly, we introduce a novel method to streamline the input schema by removing redundant parts and retaining only the parts that are truly relevant to enhance the efficiency and accuracy of the SQL generation process. Finally, we experimented with a combination of few-shot, chain-of-thought learning, and schema-enhancing methods. Through experimentation with augmented datasets, we observe encouraging initial results. Our approach outperforms the current state-of-the-art model by 23% in exact matching on the Vietnamese ViText2SQL dataset. We achieved this result with a single pretraining step and one epoch of retraining, compared to the SoTA model's 10 epochs. These findings demonstrate the effectiveness of our method and its potential for Vietnamese text-to-SQL applications.

Primary Area: unsupervised, self-supervised, semi-supervised, and supervised representation learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 11478