Granted Patent B2

US 12,353,408 · App. 17/631,576 · Granted Jul 8, 2025

Inventors: Yan Gao (Redmond, WA); Lei Fang (Redmond, WA); Dongmei Zhang (Beijing, CN); Yan Xiao (Redmond, WA); Jian-Guang Lou (Redmond, WA); Maria del Mar Gines Marin (Kirkland, WA)

Assignee: Microsoft Technology Licensing, LLC

G06F16/24522G06F16/243

Abstract

Claims (41)

1. A computer-implemented method, comprising:

receiving, at a semantic parsing unit, a first natural language query for a first data table having data in a first field of knowledge, the first natural language query comprising a plurality of words;

obtaining, via the semantic parsing unit, a first set of knowledge representations extracted from the first data table, the first set of knowledge representations describing, in a structured manner, knowledge required for understanding a natural language query for the first data table in the first field of knowledge, wherein the first set of knowledge representations includes a first knowledge type describing an entity in the first data table and at least one word in the entity, and semantic information related to the at least one word, a second knowledge type describing an attribute of entity data, and a third knowledge type describing an act related to the entity;

converting, via the semantic parsing unit, the first natural language query into a first sequence of symbols that are universal for multiple data tables corresponding to multiple different fields of knowledge and include semantic information for symbols by replacing, based on the first set of knowledge representations, the plurality of words of the first natural language query with respective symbols in a predetermined symbol table;

parsing the first sequence of symbols via the semantic parsing unit comprised of a semantic parsing model trained on multiple examples of natural language queries and corresponding knowledge representations expressed as the symbols that are universal and corresponding computer executable queries from different fields of knowledge; and

generating a first computer-executable query corresponding to the first natural language query without having to retrain the semantic parsing model for different fields of knowledge.

2. The method of claim 1 , wherein the first set of knowledge representations indicates a data column in the first data table related to the knowledge representation, and

wherein at least one word in the first natural language query is replaced with a different word based on a respective symbol in response to matching a respective knowledge representation.

3. The method of claim 2 , wherein:

the first knowledge type describes an entity or a column name of a data column in the first data table,

the second knowledge type describes an attribute of a data column in the first data table, and

the third knowledge type describes an act related to a data column in the first data table.

4. The method of claim 3 , wherein the predetermined symbol table comprises a first predetermined symbol corresponding to the first knowledge type and a symbol attribute type, and wherein the replacing the plurality of words with respective symbols in a predetermined symbol table comprises:

in response to determining that a first word of the plurality of words matches a column name of a data column in the first data table or matches the at least one word indicated by the first knowledge representation, replacing the first word with the first predetermined symbol.

5. The method of claim 3 , wherein the predetermined symbol table comprises a second predetermined symbol corresponding to the second knowledge type, and wherein the replacing the plurality of words with respective symbols in a predetermined symbol table comprises:

in response to determining that a second word of the plurality of words matches the at least one word indicated by the second knowledge representation, replacing the second word with the second predetermined symbol.

6. The method of claim 3 , wherein the predetermined symbol table comprises a third predetermined symbol corresponding to the third knowledge type, and wherein the replacing the plurality of words with respective symbols in a predetermined symbol table comprises:

in response to determining that a third word of the plurality of words matches the at least one word indicated by the third knowledge representation, replacing the third word with the third predetermined symbol.

7. The method of claim 1 , wherein the replacing the plurality of words with respective symbols in a predetermined symbol table comprises:

in response to determining that a fourth word of the plurality of symbols matches a fourth predetermined symbol in the predetermined symbol table, replacing the fourth word with the fourth predetermined symbol.

8. The method of claim 1 , wherein the generating a first computer-executable query comprises:

converting, via the semantic parsing unit, the first sequence of symbols into a plurality of logic representations by applying a set of deductive rules to the first sequence of symbols, each logic representation corresponding to a predicted semantic meaning of the first natural language query;

selecting, via the semantic parsing model, from the plurality of logic representations, a logic representation that best matches true semantics of the first natural language query via the semantic parsing model; and

generating, via the semantic parsing model, the first computer-executable query based on the selected logic representation and by replacing at least one word in the first natural language query with a word corresponding to one of the first sequence of symbols.

9. The method of claim 8 , wherein the selecting a logic representation comprises:

selecting, from the plurality of logic representations, the logic representation using the semantic parsing model.

10. The method of claim 1 , wherein the generating a first computer-executable query comprises:

generating, using the trained semantic parsing model, the first computer-executable query based on the first sequence of symbols.

11. The method of claim 10 , further comprising:

obtaining a second natural language query for a second data table and a second computer-executable query corresponding to the second natural language query;

obtaining a second set of knowledge representations associated with the second data table, the second set of knowledge representations describing, in a structured manner, knowledge required for understanding a natural language query for the second data table;

converting the second natural language query into a second sequence of symbols by replacing, based on the second set of knowledge representations, a plurality of words in the second natural language query with respective symbols in the predetermined symbol table; and

generating, at least based on the second sequence of symbols and the second computer-executable query, training data for training the neural network.

12. A computing device, comprising: a processing unit; and a memory coupled to the processing unit and storing instructions which, when executed by the processing unit, cause the device to perform acts comprising: receiving, at a semantic parsing unit, a first natural language query for a first data table having data in a first field of knowledge, the first natural language query comprising a plurality of words; obtaining, via the semantic parsing unit, a first set of knowledge representations extracted from the first data table, the first set of knowledge representations describing, in a structured manner, knowledge required for understanding a natural language query for the first data table in the first field of knowledge, wherein the first set of knowledge representations includes a first knowledge type describing an entity in the first data table and at least one word in the entity, and semantic information related to the at least one word, a second knowledge type describing an attribute of entity data, and a third knowledge type describing an act related to the entity; converting, via the semantic parsing unit, the first natural language query into a first sequence of symbols that are universal for multiple data tables corresponding to multiple different fields of knowledge and include semantic information for symbols by replacing, based on the first set of knowledge representations, the plurality of words of the first natural language query with respective symbols in a predetermined symbol table; parsing the first sequence of symbols via the semantic parsing unit comprised of a semantic parsing model trained on multiple examples of natural language queries and corresponding knowledge representations expressed as the symbols that are universal and corresponding computer executable queries from different fields of knowledge; and generating a first computer-executable query corresponding to the first natural language query without having to retrain the semantic parsing model for different fields of knowledge.

13. The device of claim 12 , wherein the first set of knowledge representations indicates a data column in the first data table related to the knowledge representation, and

wherein at least one word in the first natural language query is replaced with a different word based on a respective symbol in response to matching a respective knowledge representation.

14. The device of claim 12 , wherein the generating a first computer-executable query comprises:

converting, via the semantic parsing unit, the first sequence of symbols into a plurality of logic representations by applying a set of deductive rules to the first sequence of symbols, each logic representation corresponding to a predicted semantic meaning of the first natural language query;

selecting, via the semantic parsing model, from the plurality of logic representations, a logic representation that best matches true semantics of the first natural language query via the semantic parsing model; and

generating, via the semantic parsing model, the first computer-executable query based on the selected logic representation and by replacing at least one word in the first natural language query with a word corresponding to one of the first sequence of symbols.

15. A computer program product being tangibly stored in a computer storage medium and comprising machine-executable instructions which, when executed by a device, cause the device to perform acts comprising: receiving, at a semantic parsing unit, a first natural language query for a first data table having data in a first field of knowledge, the first natural language query comprising a plurality of words; obtaining, via the semantic parsing unit, a first set of knowledge representations extracted from the first data table, the first set of knowledge representations describing, in a structured manner, knowledge required for understanding a natural language query for the first data table in the first field of knowledge, wherein the first set of knowledge representations includes a first knowledge type describing an entity in the first data table and at least one word in the entity, and semantic information related to the at least one word, a second knowledge type describing an attribute of entity data, and a third knowledge type describing an act related to the entity; converting, via the semantic parsing unit, the first natural language query into a first sequence of symbols that are universal for multiple data tables corresponding to multiple different fields of knowledge and include semantic information for symbols by replacing, based on the first set of knowledge representations, the plurality of words of the first natural language query with respective symbols in a predetermined symbol table; parsing the first sequence of symbols via the semantic parsing unit comprised of a semantic parsing model trained on multiple examples of natural language queries and corresponding knowledge representations expressed as the symbols that are universal and corresponding computer executable queries from different fields of knowledge; and generating a first computer-executable query corresponding to the first natural language query without having to retrain the semantic parsing model for different fields of knowledge.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (26)