In work, it shows that you can ask who the winning team in college football and the right database can be automatically asked to let you know that it is actually the University of Something.
"We do not really have a way to formulate the query in the right way," Victor Zhong, one of the Salesforce researchers who operated the project, explained to News in an interview. "If I provide natural language proposals, there may be two or three ways to handle queries, we use learning reinforcement to promote the use of queries that get the same results."
You can think of how the problem of machine interpretation can quickly become massively complicated with big vocabulary. The further you can limit the number of possible translations for each missing word, the easier your problem becomes. To capitalize on this, Salesforce chooses to limit the vocabulary to terms used in database labels, words in the subject being requested and messages typically used in SQL queries.
The goal of SQL democratization is not new. Startups like ClearGraph, recently acquired by Tableau, have given it their job to open data in English instead of SQL.
"Some models show execution on the database itself," Zhong added. "But there is the possibility of a privacy issue if you raise issues about Social Security numbers."
Outside the paper itself, an important addition to Salesforce here comes in the form of a WikiSQL data set created to assist in building the model. The first HTML table is obtained from Wikipedia. This table changes the randomly generated SQL query base.
Take your time to comment on this article.