【论文笔记】SEQ2SQL: GENERATING STRUCTURED QUERIES FROM NATURAL LANGUAGE USING REINFORCEMENT LEARNING

AUGMENTED POINTER NETWORK

• 处理输入:$x=[<col>;x_1^c;x_2^c;...;x_N^c;<sql>;x^s;<question>;x^q]$

• encode: two-layer, bidirectional LSTM, the output is $h_t$

• decode: two layer, unidirectional LSTM. the output is $g_t$

• produce scaler attention: ${\alpha }_{s,t}^{ptr}=W^{ptr}tanh(U^{ptr}{g}_s+V^{ptr}{h}_t)$

• predict the next token: $softmax({\alpha }_{s,t}^{ptr})$

SEQ2SQL

• First, the network classifies an aggregation operation for the query, with the addition of a null operation that corresponds to no aggregation.
• Next, the network points to a column in the input table corresponding to the SELECT column.
• Finally, the network generates the conditions for the query using a pointer network.

$Loss=L^{agg}+L^{sel}+L^{whe}$

Aggregation Operation

• compute the scaler attention: ${\alpha }_t^{inp}=W^{inp}{h}_t^{enc}$
• softmax: ${\beta }^{inp}=softmax({\alpha }^{inp})$
• compute input representation: ${\kappa }^{agg}=\sum ^t{\beta }_t^{inp}{h}^{enc}$
• multi-layer perceptron: ${\alpha }^{agg}=W^{agg}tanh(V^{agg}{\kappa }^{agg}+b^{agg}agg)+c^{agg}$
• softmax

SELECT Column

• encode each column name with a LSTM: $e_j^c$
• compute ${\kappa }^{sel}$: method same as ${\kappa }^{agg}$
• multi-layer perceptron: ${\alpha }_j^{sel}=W^{sel}tanh(V^{sel}{\kappa }^{sel}+V^c{e}_j^c)$
• softmax

WHERE Clause

using AUGMENTED POINTER NETWORK

apply RL