SUN: Exploring Intrinsic Uncertainties in Text-to-SQL Parsers

Proceedings of the 29th International Conference on Computational Linguistics, pages 5298–5308 October 12–17, 2022. 5298 SUN: Exploring Intrinsic Uncertainties in Text-to-SQL Parsers Bowen Qin1,2, , Lihan Wang1,2, , Binyuan Hui3, , Bowen Li3, Xiangpeng Wei3 Binhua Li3, Fei Huang3, Luo Si3, Min Yang1†, Yongbin Li3† 1 Shenzhen Institute of Advanced Technology, Chinese Academy of Sciences 2 University of Chinese Academy of Sciences 3 DAMO Academy, Alibaba Group {bw.qin, lh.wang1, min.yang}@siat.ac.cn {binyuan.hby, xiangpeng.wxp, binhua.lbh, shuide.lyb}@alibaba-inc.com Abstract This paper aims to improve the performance of text-to-SQL parsing by exploring the intrinsic uncertainties in the neural network based approaches (called SUN). From the data uncertainty perspective, it is indisputable that a single SQL can be learned from multiple semantically-equivalent questions. Different from previous methods that are limited to one-to-one mapping, we propose a data uncertainty constraint to explore the under- lying complementary semantic information among multiple semantically-equivalent questions (many-to-one) and learn the robust feature representations with reduced spurious associations. In this way, we can reduce the sensitivity of the learned representations and improve the robustness of the parser. From the model uncertainty perspective, there is often structural information (dependence) among the weights of neural networks. To improve the generalizability and stability of neural text-to-SQL parsers, we propose a model uncertainty constraint to refine the query representations by enforcing the output representations of different perturbed encoding networks to be consistent with each other. Ex- tensive experiments on five benchmark datasets demonstrate that our method significantly outperforms strong competitors and achieves new state-of-the-art results. For reproducibility, we release our code and data at https: //github.com/AlibabaResearch/ DAMO-ConvAI/tree/main/sunsql. 1 Introduction Text-to-SQL parsing (Zettlemoyer and Collins, 2012; Liang et al., 2013; Zhong et al., 2017; Qin et al., 2022) aims at converting a natural lan- guage (NL) question to its corresponding structured query language (SQL) in the context of a relational database (Schema). Although relational databases Equal contribution. †Corresponding authors. can be efficiently accessed by skilled professionals via handcrafted SQLs, a natural language inter- face, whose core component relies on text-to-SQL parsing, would allow ubiquitous relational data to be accessible to a broader range of non-technical users. Therefore, text-to-SQL parsing has attracted increasing attention from both academic and indus- trial communities recently due to its broad applica- tions in question answering, conversational search interaction, and so on. Although significant efforts have been devoted to text-to-SQL parsing (Chen et al., 2021; Yu et al., 2018a; Wang et al., 2020a; Cao et al., 2021; Scholak et al., 2021) with advanced deep models and architectures by learning black-box mappings between input NL questions and output SQLs, there are still several technical challenges for accurate and robust text-to-SQL parsing. First, previous models are generally learned to fit the simplified one-to-one mapping relationship, where only one NL question is used as the input, and the appro- priate rest are ignored. However, there exists data uncertainty (Ott et al., 2018; Wei et al., 2020) in text-to-SQL parsing, i.e., one output SQL may cor- respond to multiple semantically-equivalent NL questions. At inference time, the semantic …

Content truncated to 500 words