Dialog management system

Patent metadata

Field	Value
Publication number	US-11544504-B1
Application number	US-202017022883-A
Country	US
Kind code	B1
Filing date	Sep 16, 2020
Priority date	Sep 16, 2020
Publication date	Jan 3, 2023
Grant date	Jan 3, 2023

Abstract

Techniques for determining an intent of a subsequent user input in a dialog are described. The system processes historic interaction data that is structured based on natural language understanding (NLU) hypotheses, with each NLU hypothesis being associated with one or more past user inputs received by the system, one or more sample inputs, and one or more past system responses. Based on processing of the historic interaction data and dialog data of previous turns of the dialog, the system determines candidate intents for the subsequent turn of the dialog. The system also uses context data to determine the candidate intents.

First claim

What is claimed is:

1. A computer-implemented method comprising:
receiving first interaction data corresponding to a first NLU hypothesis, the first interaction data corresponding to a plurality of users, the first interaction data including:
a first past user input, a first number of times the first past user input was received, and a second number of times the first past user input resulted in an undesired system response,
a first sample input that results in a desired system response, and
a first past system response, a third number of times the first past system response was outputted, and a fourth number of times the first past system response was an undesired system response;

receiving second interaction data corresponding to a second NLU hypothesis, the second interaction data corresponding to the plurality of users;
determining first encoded interaction data using the first interaction data;
determining second encoded interaction data using the second interaction data;
receiving dialog data corresponding to a dialog session, the dialog data including:
a first user input received from a device, and
a first system-generated response to the first user input,

determining encoded dialog data using the dialog data;
processing the first encoded interaction data, the second encoded interaction data and the encoded dialog data to determine a first plurality of NLU hypotheses corresponding to a subsequent turn of the dialog session;
receiving context data corresponding to the dialog session;
processing the encoded dialog data, the first plurality of NLU hypotheses and the context data to select a third NLU hypothesis as corresponding to the subsequent turn;
determining output data using the third NLU hypothesis; and
sending the output data to the device.

2. The computer-implemented method of claim 1 , wherein determining the first encoded interaction data comprises:
determining, using a first encoder, first encoded data corresponding to the first past user input the first number of times, and the second number of times;
receiving skill data corresponding to a skill invoked during the dialog session;
determining, using a second encoder, second encoded data corresponding to the first sample input and the skill data;
determining, using a third encoder, third encoded data corresponding to the first past system response; and
determining the first encoded interaction data using the first encoded data, the second encoded data and the third encoded data.

3. The computer-implemented method of claim 1 , further comprising:
receiving third interaction data corresponding to a first user that provided the first user input;
determining third encoded interaction data using the third interaction data;
processing the encoded dialog data and the third encoded interaction data to determine a second plurality of NLU hypotheses corresponding to the subsequent turn;
receiving fourth interaction data corresponding to a group of users including the first user;
determining fourth encoded interaction data using the fourth interaction data;
processing the encoded dialog data and the fourth encoded interaction data to determine a third plurality of NLU hypotheses corresponding to the subsequent turn;
determining a fourth plurality of NLU hypotheses using the first plurality, the second plurality and the third plurality; and
processing the fourth plurality and the context data to select the third NLU hypothesis.

4. A computer-implemented method comprising:
receiving dialog data for a dialog session, the dialog data comprising a first user input and a first system-generated response to the first user input;
receiving structured data representing first interaction data corresponding to a plurality of intents, the structured data including at least first data associated with a first intent, the first data comprising:
a first past user input,
a sample input, and
a first past system response to the first past user input;

determining encoded dialog data using the dialog data;
determining first encoded interaction data using the structured data; and
processing at least the encoded dialog data and the first encoded interaction data to determine a second intent corresponding to a predicted subsequent turn of the dialog session.

5. The computer-implemented method of claim 4 , wherein receiving the structured data comprises receiving the structured data including the first data, the first data further comprising:
a first number of times the first past user input is received from a plurality of users;
a second number of times an undesired response is outputted in response to the first past user input;
a third number of times the first past system response is outputted for the plurality of users; and
a fourth number of times the first past system response is an undesired response.

6. The computer-implemented method of claim 5 , further comprising:
determining, using a first encoder, first encoded data corresponding to the first past user input, the first number of times and the second number of times;
receiving skill data corresponding to a skill invoked during the dialog session;
determining, using a second encoder, second encoded data corresponding to the sample input and the skill data;
determining, using a third encoder, third encoded data corresponding to the first past system response; and
determining first encoded interaction data using the first encoded data, the second encoded data and the third encoded data, the first encoded interaction data corresponding to the first intent.

7. The computer-implemented method of claim 4 , wherein receiving the dialog data comprises receiving the dialog data further comprising:
first word data corresponding to a first word of the first user input for a first turn of the dialog session, and
second word data corresponding to a second word of the first system-generated response for the first turn;
wherein the first word data comprises:
first token data representing the first word;
a first identity tag representing the first word corresponds to the first user input; and
a first turn tag representing the first word corresponds to the first turn,

wherein the second word data comprises:
second token data representing the second word;
a second identity tag representing the second word corresponds to the first system-generated response; and
a second turn tag representing the second word corresponds to the first turn,

the method further comprises:
processing the first word data and the second word data using an encoder; and
determining the encoded dialog data.

8. The computer-implemented method of claim 4 , wherein receiving the dialog data comprises receiving the dialog data further comprising:
first word data representing a first word of a first turn of the dialog session, and
second word data representing a second word of a second turn of the dialog session; and
the method further comprises:
determining first turn data by processing the first word data using a first encoder;
determining second turn data by processing the second word data using a second encoder; and
determining the encoded dialog data using the first turn data and the second turn data.

9. The computer-implemented method of claim 4 , further comprising:
processing the encoded dialog data and the first encoded interaction data to determine a first plurality of intents corresponding to the predicted subsequent turn, the first interaction data representing interactions between a system and a plurality of users;
receiving second interaction data represented in the structured data, the second interaction data representing interactions between the system and a first user t

Assignees

• Amazon Tech Inc

Inventors

• Fan Xing
• Pham Hung Tuan
• Guo Chenlei
• Liu Xiaohu
• Tang Shuting

Classifications

• G06F18/2148 Primary

  characterised by the process organisation or structure, e.g. boosting cascade · CPC title

• G06F40/35

  Discourse or dialogue representation · CPC title

• G06F18/2178

  based on feedback of a supervisor · CPC title

• G06F16/90332

  Natural language query formulation or dialogue systems · CPC title

• G06K9/6257 Primary

  Physics · mapped topic