Generating conversation descriptions using neural networks

Patent metadata

Field	Value
Publication number	US-11880666-B2
Application number	US-201916517756-A
Country	US
Kind code	B2
Filing date	Jul 22, 2019
Priority date	Feb 1, 2019
Publication date	Jan 23, 2024
Grant date	Jan 23, 2024

Abstract

A description of a conversation may be generated to allow a person to understand important aspects of the conversation without needing to review the conversation. The conversation description may be generated by identifying one or more events that occurred in the conversation and then generating the description using the identified events. A set of possible events may be determined in advance for a particular application. The events may be identified by using an event neural network for each event. Each event neural network may process the messages of the conversation to generate an event score that indicates a match between the conversation and the corresponding event. The event scores may then be used to select one or more events. Message scores from the event neural network of a selected event may then be used to select one or more messages of the conversation as a rationale for the selected event.

First claim

What is claimed is:

1. A computer-implemented method for generating a description of a conversation, the method comprising:
obtaining a sequence of messages of the conversation;
computing a sequence of message embeddings by processing the sequence of messages with a message embedding neural network, wherein a message embedding is a vector in a vector space that represents a meaning of a corresponding message;
identifying one or more events from a plurality of possible events in the conversation by:
obtaining a plurality of event neural networks, wherein each event neural network corresponds to a possible event,
computing an event score for each possible event, wherein computing a first event score for a first possible event comprises processing the sequence of message embeddings with a first event neural network corresponding to the first possible event, and
selecting the one or more events for the conversation using the event scores and using a probability for a rationale, wherein the probability of the rationale includes at least one of P(rationale|action, conversation) or P(rationale|topic, conversation);

generating a text description of the conversation using the one or more identified events;
obtaining first message scores from the first event neural network corresponding to a first identified event of the one or more identified events, wherein the first message scores indicate a match between the sequence of messages and the first identified event; and
selecting, using the first message scores, one or more of the sequence of messages of the conversation as the rationale for the selection of the first identified event, wherein the selected one or more of the sequence of messages of the rationale provide an explanation of the selection of the first identified event.

2. The computer-implemented method of claim 1 , comprising:
providing the text description of the conversation to a user.

3. The computer-implemented method of claim 1 , wherein computing the sequence of message embeddings comprises:
obtaining word embeddings for words of the sequence of messages, wherein a word embedding is a vector in a vector space that represents a meaning of a corresponding word; and
processing the word embeddings with the message embedding neural network.

4. The computer-implemented method of claim 1 , wherein each message of the sequence of messages includes text indicating a type of a person who sent the message.

5. The computer-implemented method of claim 4 , wherein the type of person comprises a customer or a customer service representative.

6. The computer-implemented method of claim 1 , wherein each event neural network computes a message score for each message of the sequence of messages.

7. The computer-implemented method of claim 1 , wherein computing the first event score for the first possible event comprises:
computing action scores for a plurality of possible actions that may occur during conversations;
computing topic scores for a plurality of possible topics of conversations; and
computing the first event score using the action scores and the topic scores.

8. The computer-implemented method of claim 1 , wherein each event neural network comprises a multi-layer perceptron.

9. A system for generating a description of a conversation, the system comprising:
at least one server computer comprising at least one processor and at least one memory, the at least one server computer configured to:
obtain a sequence of messages of the conversation;
compute a sequence of message embeddings by processing the sequence of messages with a message embedding neural network, wherein a message embedding is a vector in a vector space that represents a meaning of a corresponding message;
identify one or more events from a plurality of possible events in the conversation by:
obtaining a plurality of event neural networks, wherein each event neural network corresponds to a possible event,
computing an event score for each possible event, wherein computing a first event score for a first possible event comprises processing the sequence of message embeddings with a first event neural network corresponding to the first possible event, and
selecting the one or more events for the conversation using the event scores and using a probability for a rationale, wherein the probability of the rationale includes at least one of P(rationale|action, conversation) or P(rationale|topic, conversation);

generate a text description of the conversation using the one or more identified events;
obtain first message scores from the first event neural network corresponding to a first identified event of the one or more identified events, wherein the first message scores indicate a match between the sequence of messages and the first identified event; and
select, using the first message scores, one or more of the sequence of messages of the conversation as the rationale for the selection of the first identified event,

wherein the selected one or more of the sequence of messages of the rationale provide an explanation of the selection of the first identified event.

10. The system of claim 9 , wherein the at least one server computer is configured to provide the text description of the conversation to a user.

11. The system of claim 10 , wherein:
the sequence of messages is between a customer of a company and a first customer service representative of the company; and
the user is a second customer service representative of the company.

12. The system of claim 9 , wherein each event of the plurality of possible events is associated with event description text, and generating the text description of the conversation comprises combining the event description text for each event of the one or more identified events.

13. The system of claim 12 , wherein first event description text of a first identified event comprises a slot, and the at least one server computer is configured to generate the text description of the conversation by replacing the slot with entity text from a first message of the conversation.

14. The system of claim 13 , wherein the entity text is obtained by performing named entity recognition on the sequence of messages.

15. The system of claim 9 , wherein the at least one server computer is configured to:
receive a request to view the rationale for the first identified event from a user; and
transmit the selected one or more of the sequence of messages to the user.

16. One or more non-transitory, computer-readable media comprising computer-executable instructions that, when executed, cause at least one processor to perform actions comprising:
obtaining messages of a conversation;
computing message embeddings by processing the messages with a message embedding neural network, wherein a message embedding is a vector in a vector space that represents a meaning of a corresponding message;
identifying one or more events, from a plurality of possible events in the conversation by:
obtaining a plurality of event neural networks, wherein each event neural network corresponds to a possible event,
computing an event score for each possible event, wherein computing a first event score for a first possible event comprises processing the message embeddings with a first event neural network corresponding to the first possible event, and
selecting the one or more events for the conversation using the event scores and using a probability for a rationale, wherein the probability of the rationale includes at least one of P(rationale|action, conversation) or P(rationale|topic, conversation);

gener

Assignees

• Asapp Inc

Inventors

• Yang Kevin
• Chen Howard
• Lei Tao
• Henry Shawn

Classifications

• G06N3/0442

  characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU] · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06F40/35 Primary

  Discourse or dialogue representation · CPC title

• G06F40/56

  Natural language generation · CPC title

• G06N3/044

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