Granted Patent B2

US 12,547,650 · App. 18/691,796 · Granted Feb 10, 2026

Inventors: Yongbin Li (Beijing, CN); Binhua Li (Beijing, CN); Xiang Shi (Hangzhou, CN); Ruiying Geng (Beijing, CN); Binyuan Hui (Beijing, CN); Jian Sun (Beijing, CN)

Assignee: ALIBABA DAMO (HANGZHOU) TECHNOLOGY CO., LTD.

G06F16/3329G06F16/3332G06F16/3344

Abstract

Claims (60)

1 . A human-computer dialogue method comprising, by one or more computer devices:

acquiring a user statement for a current round;

determining, from one or more candidate tables, one or more target tables that match the user statement;

parsing the user statement for the current round to obtain a first query statement, wherein parsing the user statement further comprising:

generating a sample query statement related to each candidate table in the one or more candidate tables;

converting the sample query statement into a first natural language statement;

obtaining training data according to the sample query statement related to each candidate table and the first natural language statement;

performing model training according to the training data to obtain the preset model; and

taking the one or more target tables and the user statement for the current round as inputs of a preset model to obtain the first query statement through the preset model;

querying from the one or more target tables according to the first query statement to obtain target data, wherein querying from the one or more target tables further comprising:

processing the first query statement according to a historical query statement corresponding to a user historical statement in historical dialogue data, a conversation action corresponding to the user statement for the current round, and a system state corresponding to the historical dialogue data, to obtain a second query statement; and

querying from the one or more target tables according to the second query statement to obtain the target data;

wherein:

the conversation action is obtained according to the historical query statement and the first query statement, and the conversation action comprises adding a query condition, deleting the query condition and modifying the query condition; and

the system state corresponding to the historical dialogue data is used to characterize whether a historical system reply in the historical dialogue data is a query statement;

and

generating a system reply for the current round according to the target data.

2 . The method according to claim 1 , wherein generating the system reply for the current round according to the target data comprises: generating the system reply for the current round according to the target data and the user statement for the current round.

3 . The method according to claim 1 , wherein performing the model training according to the training data comprises: training a pre-trained model according to the training data to obtain the preset model, the pre-trained model being a model trained in advance according to table information.

4 . The method according to claim 1 , wherein obtaining the training data according to the sample query statement related to each candidate table and the first natural language statement comprises: replacing a keyword in the first natural language statement with a synonym or a synonymous phrase corresponding to the keyword to obtain one or more second natural language statements corresponding to the sample query statement; and

obtaining the training data according to the sample query statement related to each candidate table, the first natural language statement and the one or more second natural language statements.

5 . An electronic device, comprising:

a memory;

a processor; and

a computer program;

wherein the computer program is stored in the memory and configured to be executed by the processor to implement the following steps;

acquiring a user statement for a current round;

determining, from one or more candidate tables, one or more target tables that match the user statement;

parsing the user statement for the current round to obtain a first query statement, wherein parsing the user statement further comprising:

generating a sample query statement related to each candidate table in the one or more candidate tables;

converting the sample query statement into a first natural language statement;

obtaining training data according to the sample query statement related to each candidate table and the first natural language statement;

performing model training according to the training data to obtain the preset model; and

taking the one or more target tables and the user statement for the current round as inputs of a preset model to obtain the first query statement through the preset model;

querying from the one or more target tables according to the first query statement to obtain target data, wherein querying from the one or more target tables further comprising:

processing the first query statement according to a historical query statement corresponding to a user historical statement in historical dialogue data, a conversation action corresponding to the user statement for the current round, and a system state corresponding to the historical dialogue data, to obtain a second query statement; and

querying from the one or more target tables according to the second query statement to obtain the target data;

wherein:

the conversation action is obtained according to the historical query statement and the first query statement, and the conversation action comprises adding a query condition, deleting the query condition and modifying the query condition; and

the system state corresponding to the historical dialogue data is used to characterize whether a historical system reply in the historical dialogue data is a query statement;

and

generating a system reply for the current round according to the target data.

6 . The electronic device according to claim 5 , wherein generating the system reply for the current round according to the target data comprises: generating the system reply for the current round according to the target data and the user statement for the current round.

7 . A non-transitory computer-readable storage medium, on which a computer program is stored, wherein the computer program, when executed by a processor, implements the following steps:

acquiring a user statement for a current round;

determining, from one or more candidate tables, one or more target tables that match the user statement;

parsing the user statement for the current round to obtain a first query statement, wherein parsing the user statement further comprising:

generating a sample query statement related to each candidate table in the one or more candidate tables;

converting the sample query statement into a first natural language statement;

obtaining training data according to the sample query statement related to each candidate table and the first natural language statement;

performing model training according to the training data to obtain the preset model; and

taking the one or more target tables and the user statement for the current round as inputs of a preset model to obtain the first query statement through the preset model;

querying from the one or more target tables according to the first query statement to obtain target data, wherein querying from the one or more target tables further comprising:

processing the first query statement according to a historical query statement corresponding to a user historical statement in historical dialogue data, a conversation action corresponding to the user statement for the current round, and a system state corresponding to the historical dialogue data, to obtain a second query statement; and

querying from the one or more target tables according to the second query statement to obtain the target data;

wherein:

the conversation action is obtained according to the historical query statement and the first query statement, and the conversation action comprises adding a query condition, deleting the query condition and modifying the query condition; and

the system state corresponding to the historical dialogue data is used to characterize whether a historical system reply in the historical dialogue data is a query statement;

and

generating a system reply for the current round according to the target data.

Assignments (1)

Priority Claims (1)

Continuity (1)

References Cited (21)