TagNet: Tag Out the Value Sequence of SQL Statement

Abstract

In order to assist person who doesn't know how to write SQL to access the data in a relation database, using a deep neural architecture to translate the natural language to SQL has recently been extensively studied. Previous work suffers from the complexity of where clause, since the number of conditions is totally random and predicting the value of condition is a sequence-to-sequence problem. We follow the slot filling idea, and introduce a model called TagNet. First of all, we innovatively propose a task attention mechanism. It takes the relativity of tasks into consideration for attention mechanism. Secondly, we use type embedding of each token of question and each column to enhance the representation for value prediction. Thirdly, in the task of predicting WHERE VALUE, we propose a tag decoder. It output a sequence of equal length compared with input. It consists of two tokens:, , indicating the corresponding token of input is whether or not a value token. We evaluate out model on WikiSQL, and compared to our baseline-SQLNet, we gain an absolute 7.6% increase on logic form accuracy and 6.3% increase on execution accuracy.