US 11,789,945 B2
Clause-wise text-to-SQL generation
Dongjun Lee, Seoul (KR)
Assigned to SAP SE, Walldorf (DE)
Filed by SAP SE, Walldorf (DE)
Filed on Dec. 6, 2019, as Appl. No. 16/705,570.
Claims priority of provisional application 62/835,588, filed on Apr. 18, 2019.
Prior Publication US 2020/0334252 A1, Oct. 22, 2020
Int. Cl. G06F 16/2452 (2019.01); G06N 3/08 (2023.01); G06F 16/242 (2019.01)
CPC G06F 16/24522 (2019.01) [G06F 16/243 (2019.01); G06N 3/08 (2013.01)] 20 Claims
OG exemplary drawing
 
1. A computing system comprising:
a processor configured to
receive a natural language input,
determine, via execution of a long short term memory (LSTM) neural network, a skeleton data structure of a structured query language (SQL) operation which includes an unfilled SQL clause based on text included in the natural language input,
identify, via execution of the LSTM neural network, a missing column name exists in the unfilled SQL clause and insert an empty slot into the unfilled SQL clause within the skeleton data structure for the missing column name,
convert a plurality of column names from a database into column name vectors via an encoder of an encoder-decoder neural network,
input the column name vectors and a vector of the natural language input into a decoder of the encoder-decoder neural network,
select a column name from among the plurality of column names which most closely matches one or more words in the natural language input via a decoder of the encoder-decoder neural network based on a comparison of the column name vectors output from the encoder and the natural language input vector; and
fill in the empty slot corresponding to the missing column name in skeleton data structure with the selected column name from the decoder; and
a memory configured to store the filled-in skeleton data structure of the SQL operation.