Granted Patent B2

US 12,530,344 · App. 18/725,248 · Granted Jan 20, 2026

Inventors: Wook Shin Han (Pohang-si, KR); Hyuk Kyu Kang (Pohang-si, KR); Hyeon Ji Kim (Ulsan, KR)

Assignee: POSTECH RESEARCH AND BUSINESS DEVELOPMENT FOUNDATION

G06F16/242G06F16/211

Abstract

Claims (105)

1 . A method of augmenting training datasets used for training a predetermined neural network in a natural language processing system, the natural language processing system including a translator for translating a natural language query into a structured query language (SQL) query based on the predetermined neural network, a data augmentation engine augmenting the training datasets for the translator, and a training engine training the translator using the augmented training datasets, the method comprising:

determining, with the data augmentation engine, a first initial training dataset including a first training natural language query, database information of a database, and a first training SQL query corresponding to the first training natural language query;

determining, with the data augmentation engine, a first schema modification operation from among predetermined schema modification operations;

generating, with the data augmentation engine, first new database information of a first new database having a different schema than the database by applying the first schema modification operation to the database;

determining, with the data augmentation engine, whether the first schema modification operation transforms the first training SQL query;

when the first schema modification operation is determined to transform the first training SQL query, generating, with the data augmentation engine, a first new training SQL query for the first new database by applying a first SQL query synchronization operation to the first training SQL query, the first SQL query synchronization operation being determined based on predetermined SQL query synchronization operation definition information corresponding to the first schema modification operation;

when the first schema modification operation is determined not to transform the first training SQL query, determining, with the data augmentation engine, the first training SQL query as the first new training SQL query for the first new database;

determining, with the data augmentation engine, a first new training dataset including the first training natural language query, the first new database information, and the generated or determined first new training SQL query; and

providing, with the data augmentation engine, the first new training dataset as input to the training engine.

2 . The method of claim 1 , further comprising:

generating, with the data augmentation engine, second new database information of a second new database having a different schema than the database by applying a second schema modification operation, determined from among the predetermined schema modification operations, to the database;

generating, with the data augmentation engine, a second new training SQL query for the second new database by applying a second SQL query synchronization operation corresponding to the second schema modification operation to the first training SQL query; and

determining, with the data augmentation engine, a second new training dataset including the first training natural language query, the second new database information, and the second new training SQL query.

3 . The method of claim 2 , further comprising:

determining, with the data augmentation engine, a second initial training dataset including a second training natural language query, the database information, and a second training SQL query corresponding to the second training natural language query;

generating, with the data augmentation engine, third new database information of a third new database having a different schema than the database by applying the first schema modification operation to the database;

generating, with the data augmentation engine, a third new training SQL query for the third new database by applying the first SQL query synchronization operation to the second training SQL query; and

determining, with the data augmentation engine, a third new training dataset including the second training natural language query, the third new database information, and the third new training SQL query.

4 . The method of claim 3 , further comprising:

generating, with the data augmentation engine, fourth new database information of a fourth new database having a different schema than the first new database by applying the first schema modification operation to the first new database;

generating, with the data augmentation engine, a fourth new training SQL query for the fourth new database by applying the first SQL query synchronization operation to the first new training SQL query; and

determining, with the data augmentation engine, a fourth new training dataset including the first training natural language query, the fourth new database information, and the fourth new training SQL query.

5 . The method of claim 1 , further comprising:

determining, with the data augmentation engine, a second initial training dataset including a second training natural language query, the database information, and a second training SQL query corresponding to the second training natural language query;

generating, with the data augmentation engine, second new database information of a second new database having a different schema than the database by applying the first schema modification operation to the database;

generating, with the data augmentation engine, a second new training SQL query for the second new database by applying the first SQL query synchronization operation corresponding to the first schema modification operation to the second training SQL query; and

determining, with the data augmentation engine, a second new training dataset including the second training natural language query, the second new database information, and the second new training SQL query.

6 . The method of claim 1 , wherein the first schema modification operation includes at least one of a first modification operation of changing a schema structure of the database or a second modification operation of changing a name of a schema element of the database.

7 . A method of training a predetermined neural network in a natural language processing system, the natural language processing system including a translator for translating a natural language query into a structured query language (SQL) query based on the predetermined neural network, a data augmentation engine augmenting training datasets for the translator, and a training engine training the translator using the augmented training datasets, the method comprising:

determining, with the data augmentation engine, a first initial training dataset including a first training natural language query, database information of a database, and a first training SQL query corresponding to the first training natural language query;

determining, with the data augmentation engine, a first schema modification operation from among predetermined schema modification operations;

generating, with the data augmentation engine, first new database information of a first new database having a different schema than the database by applying the first schema modification operation to the database;

determining, with the data augmentation engine, whether the first schema modification operation transforms the first training SQL query;

when the first schema modification operation is determined to transform the first training SQL query, generating, with the data augmentation engine, a first new training SQL query for the first new database by applying a first SQL query synchronization operation to the first training SQL query, the first SQL query synchronization operation being determined based on predetermined SQL query synchronization operation definition information corresponding to the first schema modification operation;

when the first schema modification operation is determined not to transform the first training SQL query, determining, with the data augmentation engine, the first training SQL query as the first new training SQL query for the first new database;

determining, with the data augmentation engine, a first new training dataset including the first training natural language query, the first new database information, and the generated or determined first new training SQL query;

training, with the training engine, the neural network using the first new training dataset received from the data augmentation engine to optimize parameters of the translator; and

providing, with the training engine, the parameters output by the trained neural network and translation reference data generated during training of the neural network for reference in translation to the translator.

8 . The method of claim 7 , further comprising:

generating, with the data augmentation engine, second new database information of a second new database having a different schema than the database by applying a second schema modification operation, determined from among the predetermined schema modification operations, to the database;

generating, with the data augmentation engine, a second new training SQL query for the second new database by applying a second SQL query synchronization operation corresponding to the second schema modification operation to the first training SQL query; and

determining, with the data augmentation engine, a second new training dataset including the first training natural language query, the second new database information, and the second new training SQL query,

wherein the training of the neural network is performed using the first and second new training datasets.

9 . The method of claim 8 , further comprising:

determining, with the data augmentation engine, a second initial training dataset including a second training natural language query, the database information, and a second training SQL query corresponding to the second training natural language query;

generating, with the data augmentation engine, third new database information of a third new database having a different schema than the database by applying the first schema modification operation to the database;

generating, with the data augmentation engine, a third new training SQL query for the third new database by applying the first SQL query synchronization operation to the second training SQL query; and

determining, with the data augmentation engine, a third new training dataset including the second training natural language query, the third new database information, and the third new training SQL query,

wherein the training of the neural network is performed using the first to third new training datasets.

10 . The method of claim 9 , further comprising:

generating, with the data augmentation engine, fourth new database information of a fourth new database having a different schema than the first new database by applying the first schema modification operation to the first new database;

generating, with the data augmentation engine, a fourth new training SQL query for the fourth new database by applying the first SQL query synchronization operation to the first new training SQL query; and

determining, with the data augmentation engine, a fourth new training dataset including the first training natural language query, the fourth new database information, and the fourth new training SQL query,

wherein the training of the neural network is performed using the first to fourth new training datasets.

11 . The method of claim 7 , further comprising:

determining, with the data augmentation engine, a second initial training dataset including a second training natural language query, the database information, and a second training SQL query corresponding to the second training natural language query;

generating, with the data augmentation engine, second new database information of a second new database having a different schema than the database by applying the first schema modification operation to the database;

generating, with the data augmentation engine, a second new training SQL query for the second new database by applying the first SQL query synchronization operation corresponding to the first schema modification operation to the second training SQL query; and

determining, with the data augmentation engine, a second new training dataset including the second training natural language query, the second new database information, and the second new training SQL query,

wherein the training of the neural network is performed using the first and second new training datasets.

12 . The method of claim 7 , wherein the first schema modification operation includes at least one of a first modification operation of changing a schema structure of the database or a second modification operation of changing a name of a schema element of the database.

13 . An information search device for providing a search result corresponding to a natural language query, the information search device comprising:

a memory configured to store program commands; and

a processor connected to the memory and configured to execute the program commands stored in the memory,

wherein, when the program commands are executed by the processor, the program commands cause the processor to perform the operations of:

training, with a training engine, a predetermined neural network;

receiving the natural language query from a user;

translating, with a translator, the natural language query into a structured query language (SQL) query based on the predetermined neural network;

acquiring, with a relational database management system, the search result corresponding to the natural language query using the SQL query; and

providing the search result to the user,

wherein the program commands causing the processor to perform the operation of training the predetermined neural network cause the processor to perform the operations of:

determining, with the data augmentation engine, a first initial training dataset including a first training natural language query, database information of a database, and a first training SQL query corresponding to the first training natural language query;

determining, with the data augmentation engine, a first schema modification operation from among predetermined schema modification operations;

generating, with the data augmentation engine, first new database information of a first new database having a different schema than the database by applying the first schema modification operation to the database;

determining, with the data augmentation engine, whether the first schema modification operation transforms the first training SQL query;

when the first schema modification operation is determined to transform the first training SQL query, generating, with the data augmentation engine, a first new training SQL query for the first new database by applying a first SQL query synchronization operation to the first training SQL query, the first SQL query synchronization operation being determined based on predetermined SQL query synchronization operation definition information corresponding to the first schema modification operation;

when the first schema modification operation is determined not to transform the first training SQL query, determining, with the data augmentation engine, the first training SQL query as the first new training SQL query for the first new database;

determining, with the data augmentation engine, a first new training dataset including the first training natural language query, the first new database information, and the generated or determined first new training SQL query;

training, with the training engine, the neural network using the first new training dataset received from the data augmentation engine to optimize parameters of the translator;

providing, with the training engine, the parameters output by the trained neural network and translation reference data generated during training of the neural network for reference in translation to the translator; and

wherein the program commands, when executed by the processor, further cause the processor, when translating, with the translator, the natural language query into the structured query language (SQL) query based on the predetermined neural network, to translate the natural language query into the SQL query using the parameters and the translation reference data.

14 . The information search device of claim 13 , wherein the program commands causing the processor to perform the operation of training the neural network cause the processor to further perform the operations of:

generating, with the data augmentation engine, second new database information of a second new database having a different schema than the database by applying a second schema modification operation, determined from among the predetermined schema modification operations, to the database;

generating, with the data augmentation engine, a second new training SQL query for the second new database by applying a second SQL query synchronization operation corresponding to the second schema modification operation to the first training SQL query; and

determining, with the data augmentation engine, a second new training dataset including the first training natural language query, the second new database information, and the second new training SQL query,

wherein the training of the neural network is performed using the first and second new training datasets.

15 . The information search device of claim 14 , wherein the program commands causing the processor to perform the operation of training the neural network cause the processor to further perform the operations of:

determining, with the data augmentation engine, a second initial training dataset including a second training natural language query, the database information, and a second training SQL query corresponding to the second training natural language query;

generating, with the data augmentation engine, third new database information of a third new database having a different schema than the database by applying the first schema modification operation to the database;

generating, with the data augmentation engine, a third new training SQL query for the third new database by applying the first SQL query synchronization operation to the second training SQL query; and

determining, with the data augmentation engine, a third new training dataset including the second training natural language query, the third new database information, and the third new training SQL query,

wherein the training of the neural network is performed using the first to third new training datasets.

16 . The information search device of claim 14 , wherein the program commands causing the processor to perform the operation of training the neural network cause the processor to further perform the operations of:

generating, with the data augmentation engine, fourth new database information of a fourth new database having a different schema than the first new database by applying the first schema modification operation to the first new database;

generating, with the data augmentation engine, a fourth new training SQL query for the fourth new database by applying the first SQL query synchronization operation to the first new training SQL query; and

determining, with the data augmentation engine, a fourth new training dataset including the first training natural language query, the fourth new database information, and the fourth new training SQL query,

wherein the training of the neural network is performed using the first to fourth new training datasets.

17 . The information search device of claim 13 , wherein the program commands causing the processor to perform the operation of training the neural network cause the processor to further perform the operations of:

determining, with the data augmentation engine, a second initial training dataset including a second training natural language query, the database information, and a second training SQL query corresponding to the second training natural language query;

generating, with the data augmentation engine, second new database information of a second new database having a different schema than the database by applying the first schema modification operation to the database;

generating, with the data augmentation engine, a second new training SQL query for the second new database by applying the first SQL query synchronization operation corresponding to the first schema modification operation to the second training SQL query; and

determining, with the data augmentation engine, a second new training dataset including the second training natural language query, the second new database information, and the second new training SQL query,

wherein the training of the neural network is performed using the first and second new training datasets.

18 . The information search device of claim 13 , wherein the first schema modification operation includes at least one of a first modification operation of changing a schema structure of the database or a second modification operation of changing a name of a schema element of the database, and

when the first schema modification operation only includes the first modification operation, the first SQL query synchronization operation is a null operation.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (14)