Fine-tuning text-to-SQL models with reinforcement-learning training objectives

Text-to-SQL is an important natural language processing task that helps users automatically convert natural language queries into formal SQL code. While transformer-based models have pushed text-to-SQL to unprecedented accuracy levels in recent years, such performance is confined to models of very large size that can only be run in specialised clouds. For this reason, in this paper we explore the use of reinforcement learning to improve the performance of models of more conservative size, which can fit within standard user hardware. As reinforcement learning reward, we propose a novel function which better aligns with the text-to-SQL evaluation metrics, applied in conjunction with two strong policy gradient algorithms, REINFORCE and RELAX. Our experimental results over the popular Spider benchmark show that the proposed approach has been able to outperform a conventionally-trained T5 Small baseline by 6.6 pp (percentage points) of exact-set-match accuracy and 4.6 pp of execution accuracy, and a T5 Base baseline by 2.0 pp and 1.9 pp, respectively. The proposed model has also achieved a remarkable comparative performance against ChatGPT instances. • Tackles the useful text-to-SQL task with models of compact size. • Leverages reinforcement learning training objectives to improve the models’ performance. • Compares with conventional training objectives and a popular policy gradient algorithm. • Results on the Spider dataset show that the proposed approach has outperformed the compared models. • Performs strongly also against zero-shot and one-shot set-ups of GPT-3.5 and GPT-4.