Their new paper, Seq2SQL: Creating Structured Queries from Natural Language applying Reinforcement Learning, makes on the sequence to sequence models typically used in machine translation. A reinforcement training twist enabled the team to obtain assuring results translating natural language database requests into SQL.
In work, this suggests that you could just ask who the winningest team in college soccer is and a proper database could be automatically asked to tell you that it is, in fact, the University of Something.
“We don’t really have just one way of formulating a query the correct way,” Victor Zhong, one of the Salesforce researchers who operated on the project, explained to News in an interview. “If I give a natural language proposal, there might be two or three ways to address the query. We use reinforcement learning to promote the use of queries that get the same result.”
You can think how machine interpretation problems can quickly become massively complicated with large vocabularies. The further you can limit the number of possible translations for each lost word, the easier your problem becomes. To this avail, Salesforce opted to restrict its vocabulary to terms used in database labels, the words in the subject being asked and the messages typically used in SQL queries.
The purpose of democratizing SQL isn’t new. Startups like ClearGraph, which was newly acquired by Tableau, have given it their job to open up data with English rather than SQL.
“Some models show the execution on a database itself,” added Zhong. “But there are possible privacy concerns if you’re asking an issue about Social Security numbers.”
Outside of the paper itself, Salesforce’s important addition here comes in the form of the WikiSQL data set is created to aid in building its model. First HTML tables were obtained from Wikipedia. These tables converted the basis for randomly generated SQL queries.
Take your time to comment on this article.