Research on BERT-based Text2SQL Multi-task Learning
Liu Xusheng, Yeteng An, Jingxian Lv, Huimin Zhang, Zhang Yumeng, Min Li, Wei Zhao, Han Wei, Sun Liangfei, Huiqin Li · 2023
With the continuous development of computer technology, the Internet has entered the era of big data, which has spawned massive data in various fields and stored it in the database in a structured form. As an important national infrastructure, the power system is also constantly generating various types of data, and has become an important resource for daily operation and maintenance. During the interaction with the database, business personnel need to manually write SQL statements, which poses a challenge to a large number of nonprofessionals; At the same time, for experienced professionals, the conditions of SQL statements are complex and changeable, which is easy to make mistakes, thus affecting the utilization of data and the efficiency of queries. Therefore, how to fully explore the relationship between natural language and SQLstatements and apply them to power operation and maintenance is an important research topic. This paper studies the method of transforming natural language into SQL statement according to the current situation of power field. In terms of solution strategy, this paper divides the problem into several subproblems and combines BERT to carry out multi task learning. The experimental results show that this method is effective on Text2SQL tasks.