[4Rin1] Interactive 2

[4Rin1-09] A Study on Text-to-SQL with Data Values

〇Soichiro Kaku¹, Kyosuke Nishida¹, Junji Tomita¹ (1.Nippon Telegraph and Telephone Corporation)

Keywords:Text-to-SQL

Text-to-SQL, which converts a natural language question for a database (DB) into an SQL query, has been attracting attention. In previous techniques, a natural language question and a DB schema are used as inputs to output an SQL query. However, they determine which columns in a DB are used in an SQL query without considering the data values. Since there is a DB consisting of tens of thousands of words, it is difficult to handle all data values in all columns simultaneously with a DB schema. In this study, we define a sub- task that determines whether or not a DB column is used in an SQL query to get the answer to a question using data values. Experiments on the Spider data set showed that considering data values with a pre-trained language model was effective in determining the columns of a DB used in an SQL query. The column determination model learned in this task can be combined with a text-to-SQL model.