Lifting the Answer: Reranking Candidates on Data Augmented Text-to-SQL

Guang-Yu Cheng, Shuo Yu, Wen-Bin Jiang, Shuai Ma, Xiang Ao, Qing He

Abstract: The text-to-SQL task is a realistic and challenging job, which aims to translate natural language questions into corresponding SQL queries. When attempting to generate SQL queries in the text-to-SQL task, prevailing semantic parsing models employ beam search to generate several candidates. Based on our pilot study, we observe that the gold-standard SQL answer may exist in the n-best candidate list produced by the decoder, rather than the first. Hence, we aim to lift the correct answer given several candidates generated by semantic parsing models for text-to-SQL. To this end, we propose a reranking module that reorders the n-best list of candidate SQL queries by pair-wise hinge loss. Meanwhile, a data augmentation module is leveraged to enrich the inadequate training instances for providing candidates with better qualities. These two modules can be easily grafted onto text-to-SQL backbone networks, and extensive experiments on the cross-domain text-to-SQL benchmark Spider demonstrate that our method achieves 73.8% in accuracy on the Spider dataset, surpassing the base model by up to 3.5%.

External IDs:dblp:conf/icmlc/ChengYJMAH23