DNN for translating NLQ to correspontin SQL

Distinct Characteristics

• Uses policy based reinforcement learning to generate the conditions of SQL



• Leverages the structure of SQL to prune the output space

Augmented Pointer Network

Pointer Networks are just a variation of seq2seq model with attention

Instead of translating one sequence into another, they yield a succession of pointers to the elements of the input series.

Augmented Pointer Network allows to look back and forth over input. Meaning it has access to encoder states from each step, not just the last one

The augmented pointer network generates the SQL query token-by-token by selecting from an input sequence. In our case, the input sequence is the concatenation of the column names, required for the selection column and the condition columns of the query, the question, required for the conditions of the query, and the limited vocabulary of the SQL language such as SELECT, COUNT etc.

The decoder network uses a two layer, unidirectional LSTM that computes a scaler attention score.