Dialog agents with two-sided modeling

Patent metadata

Field	Value
Publication number	US-12493791-B2
Application number	US-202217866263-A
Country	US
Kind code	B2
Filing date	Jul 15, 2022
Priority date	Aug 4, 2021
Publication date	Dec 9, 2025
Grant date	Dec 9, 2025

Abstract

A central learning model is deployed as a user model and as an assistant model. Sensitive information utterances from a corpus of previously stored conversation language corresponding to user queries and chat agent responses thereto are used to train the user model to become an updated user model and to train the assistant model to become an updated assistant model, respectively. The user model provides user contexts corresponding to user queries to the assistant model and the assistant model provides assistant contexts corresponding to chat agent responses to the user model. During training, the user model does not provide plain-text queries to the assistant model and the assistant model does not provide plain-text responses to the user model. The updated assistant model may facilitate a federated training process produce an updated central model. An updated central model may be used to provide real-time chat agent responses to live user queries.

First claim

What is claimed is:

1 . A method, comprising:
receiving, from a first initial learning model executing on a first computing device comprising at least one processor, first context information representative of a first context that corresponds to sensitive information; transmitting, by the first computing device to a second computing device comprising a central learning model, the first context information; receiving, from the second computing device, a second initial learning model, wherein the second initial learning model resulted from the first context information having been input to the central learning model; inputting the first context information to the second initial learning model executing on the first computing device; and determining, using the second initial learning model, reply information responsive to the sensitive information based on the first context information, wherein the first context information is derived from the sensitive information, and wherein the first context information does not comprise the sensitive information.

2 . The method of claim 1 , further comprising:
determining, by using the second initial learning model, updated context information based on the first context information and second context information representative of a second context that corresponds to the reply information; and transmitting the updated context information to the first initial learning model from the second initial learning model.

3 . The method of claim 1 , further comprising training, by the computing device, the second initial learning model based on the reply information to result in an updated second learning model.

4 . The method of claim 3 , wherein the sensitive information input to the first initial learning model is not used to train the second initial learning model that results in the updated second learning model.

5 . The method of claim 1 , wherein the second initial learning model does not receive the sensitive information.

6 . The method of claim 1 , wherein the first initial learning model and the second initial learning model comprise a pre-trained language model.

7 . The method of claim 6 , wherein the pre-trained language model comprises a generative pre-trained transformer model.

8 . The method of claim 1 , further comprising training, by the second computing device, the central learning model according to a federated learning model to result in an updated central model based on different reply information that was generated responsive to different contexts, which correspond to different conversation dialogs between different user learning models and different assistant learning models, received from the different assistant learning models.

9 . The method of claim 8 , further comprising:
determining, with a central computing device of a central computing system comprising a processor, a response to a query received from a user device that is configured to present a dialog agent application interface of the user device; and transmitting, with the central computing system, the response to the dialog agent application interface of the user device; wherein the central computing device uses the updated central model to determine the response to the query.

10 . The method of claim 8 , further comprising:
determining a response to a query input to a dialog agent of a user device that comprises a processor and that is configured to present a dialog agent application interface of the user device; wherein the user device uses the updated central model to determine the response to the query.

11 . A computing system, comprising at least one processor configured to:
receive, from a first initial learning model executing on a computing device, first context information representative of a first context that corresponds to sensitive information; transmit, to a central server comprising a central learning model, the first context information; receive a second initial learning model that was directed by the central server to the computing device, wherein the second initial learning model results from the first context information being input to the central learning model; input the first context information to the second initial learning model executing on the computing device; and determine, using the second initial learning model, reply information responsive to the sensitive information based on the first context information, wherein the first context information is derived from the sensitive information and wherein the first context information does not comprise the sensitive information.

12 . The computing system of claim 11 , wherein the at least one processor is further configured to:
determine, by using the second initial learning model, updated context information based on the first context information and second context information representative of a second context that corresponds to the reply information; and transmit the updated context information to the first initial learning model from the second initial learning model.

13 . The computing system of claim 11 , wherein the at least one processor is further configured to train the second initial learning model based on the reply information to result in an updated second learning model.

14 . The computing system of claim 13 , wherein the sensitive information input to the first initial learning model is not used in the training of the learning model that results in the updated second learning model.

15 . The computing system of claim 11 , wherein the first initial training model and the second initial training model comprise a pre-trained central language model.

16 . The computing system of claim 13 , wherein parameters corresponding to the updated second learning model are combined with parameters from other models to result in an updated central learning model.

17 . A non-transitory machine-readable medium, comprising executable instructions that, when executed by at least one processor of a computing device that comprises access to a first initial learning model and to a second initial learning model, facilitate performance of operations, comprising:
inputting first sensitive information to the first initial learning model; determining, with the first initial learning model, first context information that corresponds to the first sensitive information; transmitting the first context information to the second initial learning model; determining, with the second initial learning model, reply language information responsive to the first context information, and first updated context information based on the first context information and based on second context information that corresponds to the reply language information; transmitting the first updated context information to the first initial learning model; inputting second sensitive information responsive to the first updated context information to the first initial learning model; determining, with the first initial learning model, third context information that corresponds to the second sensitive information and the first updated context information; determining second updated context information based on the first updated context information and the third context information; and transmitting the second updated context information to the second initial learning model, wherein the first context information does not comprise the first sensitive information, wherein the third context information does not comprise the first sensitive information or the second sensitive inf

Assignees

• Univ Hong Kong Science & Tech

Inventors

• Song Yangqiu
• Li Haoran
• Su Ying
• Hu Qi
• Bai Jiaxin
• Jin Yilun

Classifications

• G06F40/35 Primary

  Discourse or dialogue representation · CPC title

• G06F16/3329

  Natural language query formulation · CPC title

• G06N3/0455

  Auto-encoder networks; Encoder-decoder networks · CPC title

• G06N3/044

  Recurrent networks, e.g. Hopfield networks · CPC title

• G06N3/084

  Backpropagation, e.g. using gradient descent · CPC title