结合数据库结构及内容的问句理解方法研究  

Research on Question Understanding Method Combining Database Structure and Content

作　　者：袁志祥 任冬冬 洪旭东 孙国华 YUAN Zhixiang;REN Dongdong;HONG Xudong;SUN Guohua(School of Computer Science and Technology,Anhui University of Technology,Maanshan,Anhui 243000,China)

机构地区：[1]安徽工业大学计算机科学与技术学院,安徽马鞍山243000

出　　处：《计算机工程》2021年第3期71-76,82,共7页Computer Engineering

基　　金：国家重点研发计划(2016YFF020440508)。

摘　　要：问句理解是模型将自然语言问句转换成SQL的重要基础。目前多数利用深度学习的模型仅是通过数据库结构,未结合数据库内容充分理解问句生成SQL查询。在SQLova模型的基础上,提出一种基于表结构和内容的问句理解方法。利用表结构和表内容关注机制获得问句更准确的语义表达式,通过子类分类任务填充SQL草图完成SQL查询。在阿里云首届中文NL2SQL挑战赛发布的中文数据集上进行测试,结果表明,结合数据库结构与内容的问句理解方法取得78%的准确率,比不结合表内容的模型高出1.8%,在WikiSQL数据集上比SQLova准确率高出1.4%,可以有效提高生成SQL查询的准确率。Question understanding is an important basis for the model to transform natural language questions into SQL.At present,most of the models using deep learning generate database queries through only the database structure and do not fully understand the question in combination with the database contents.To address this problem,based on the SQLova model,this paper proposes a question understanding method combining database structure and content.This method uses the attention mechanisms for table structure and table content to obtain more accurate semantic expressions of the question.The sub-class classification task is used to fill the SQL sketch to complete the SQL query.The model is tested on a Chinese data set released on the first Chinese NL2SQL Challenge of Alibaba Cloud.Test results show that the proposed question understanding method combining database structure and content achieves an accuracy rate of 78%,which is 1.8%higher than that of the model without table contents.When the method is tested on the WikiSQL dataset,its accuracy rate is 1.4%higher than that of SQLova.The experiment results show that the proposed method can effectively improve the accuracy of the generated SQL queries.

关 键 词：表结构 表内容 问句理解 槽填充 SQL查询 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]