Customized message suggestion with user embedding vectors

Patent metadata

Field	Value
Publication number	US-2021126881-A1
Application number	US-201916663872-A
Country	US
Kind code	A1
Filing date	Oct 25, 2019
Priority date	Oct 25, 2019
Publication date	Apr 29, 2021
Grant date	—

Abstract

A message may be suggested to a user participating in a conversation using one or more neural networks where the suggested message is adapted to the preferences or communication style of the user. The suggested message may be adapted to the user with a user embedding vector that represents the preferences or communication style of the user in a vector space. To suggest a message to the user, a conversation feature vector may be computed by processing the text the conversation with a neural network. A context score may be computed for one or more designated messages, where the context score is computed by processing the user embedding vector, the conversation feature vector, and a designated message feature vector with a neural network. A designated message may be selected as a suggested message for the user using the context scores. The suggestion may then presented to the user.

First claim

What is claimed is:

1 . A computer-implemented method, comprising:
obtaining a user embedding vector corresponding to a user, wherein the user embedding vector represents the user in a first vector space; receiving text of a conversation with the user; computing a conversation feature vector using the text of the conversation and a context encoder neural network, wherein the conversation feature vector represents the conversation in a second vector space; obtaining a set of designated messages, wherein each designated message is associated with a corresponding designated message feature vector; computing a first context score for a first designated message of the set of designated messages by processing the user embedding, the conversation feature vector, and a first designated message feature vector with a context scoring neural network; computing a second context score for a second designated message of the set of designated messages by processing the conversation feature vector, a second designated message feature vector, and the user embedding with the context scoring neural network; computing a third context score for a third designated message of the set of designated messages by processing the conversation feature vector, a third designated message feature vector, and the user embedding with the context scoring neural network; selecting the first designated message using the first context score, the second context score, and the third context score; presenting the first designated message as a suggested message to the user; receiving a selection of the first designated message from the user; and transmitting the first designated message in the conversation as a message from the user.

2 . The computer-implemented method of claim 1 , wherein computing the conversation feature vector comprises:
obtaining word embeddings for the text of the conversation, wherein a word embedding represents a corresponding word in a third vector space; and processing the word embeddings with the context encoder neural network.

3 . The computer-implemented method of claim 1 , comprising:
selecting the second designated message using the first context score, the second context score, and the third context score; and presenting the second designated message as a suggested message to the user.

4 . The computer-implemented method of claim 1 , comprising:
obtaining text encodings for text of the first designated message; and computing the first designated message feature vector by processing the text encodings for the text of the first designated message with a response encoder neural network.

5 . The computer-implemented method of claim 1 , comprising:
obtaining input text entered by the user; computing a first input text score for the first designated message; computing a second input text score for the second designated message; computing a third input text score for the third designated message; and selecting the first designated message using the first input text score, the second input text score, and the third input text score.

6 . The computer-implemented method of claim 1 , comprising:
obtaining input text entered by the user; obtaining text encodings for the input text; computing an input text feature vector by processing the text encodings of the input text with a response encoder neural network; computing a first semantic score for the first designated message using the input text feature vector and the first designated message feature vector; computing a second semantic score for the second designated message using the input text feature vector and the second designated message feature vector; computing a third semantic score for the third designated message using the input text feature vector and the third designated message feature vector; and selecting the first designated message using the first semantic score, the second semantic score, and the third semantic score.

7 . The computer-implemented method of claim 6 , wherein computing the first semantic score comprises computing a distance or a cosine similarity between the input text feature vector and the first designated message feature vector.

8 . The computer-implemented method of claim 1 , wherein the context scoring neural network comprises a multi-layer perceptron.

9 . A 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 user embedding vector corresponding to a user, wherein the user embedding vector represents the user in a first vector space;
receive text of a conversation with the user;
compute a conversation feature vector using the text of the conversation and a context encoder neural network, wherein the conversation feature vector represents the conversation in a second vector space;
obtain a set of designated messages, wherein each designated message is associated with a corresponding designated message feature vector;
compute a first context score for a first designated message of the set of designated messages by processing the user embedding, the conversation feature vector, and a first designated message feature vector with a context scoring neural network;
compute a second context score for a second designated message of the set of designated messages by processing the conversation feature vector, a second designated message feature vector, and the user embedding with the context scoring neural network;
select the first designated message using the first context score and the second context score; and
present the first designated message as a suggested message to the user.

10 . The system of claim 9 , wherein the user embedding vector is trained using messages sent by the user and represents a conversational style or conversational preferences of the user.

11 . The system of claim 9 , wherein the user is a customer support agent of a company and the system is implemented by a third-party company providing services to the company.

12 . The system of claim 9 , wherein the at least one server computer is configured to:
obtain input text entered by the user; obtain a first prefix score for the first designated message by comparing the input text to a prefix of the first designated message; obtain a second prefix score for the second designated message by comparing the input text to a prefix of the second designated message; select the first designated message using the first prefix score and the second prefix score.

13 . The system of claim 12 , wherein a prefix score has a first value if the input text is not a prefix of a corresponding designated message and has a second value if the input text is a prefix of the corresponding designated message.

14 . The system of claim 9 , wherein the at least one server computer is configured to:
compute a first semantic score for the first designated message; compute a second semantic score for the second designated message; compute a first prefix score for the first designated message; compute a second prefix score for the second designated message; select the first designated message using the first context score, the second context score, the first prefix score, and the second prefix score.

15 . The system of claim 9 , wherein the context encoder neural network comprises a recurrent neural network.

16 . The system of claim 9 , wherein the at least one server computer is configured to compute the first context score by processing conversation metadata with the context scoring neu

Assignees

• Asapp Inc

Inventors

• Ball Kelsey Taylor
• Lei Tao
• Fox Christopher David
• Hackman Joseph Ellsworth

Classifications

• G06N3/044

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

• G06N3/0455

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

• G06N3/0499

  Feedforward networks · CPC title

• G06N3/09

  Supervised learning · CPC title

• G06N3/0442

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