Maintaining quality of customer support messages

• US 10,489,792 B2
• Filed: 01/05/2018
• Issued: 11/26/2019
• Est. Priority Date: 01/05/2018
• Status: Active Grant

• Current Assignee: Asapp, Inc.
• Original Assignee: Asapp, Inc.
• Inventors: Hackman, Joseph Ellsworth, Henry, Shawn, Hampton, Alan Nicolas, Lei, Tao
• Primary Examiner(s): Spooner, Lamont M
• Application Number:
  US15/863,225
• Publication Number:
  US 20190213601A1
• Time in Patent Office: 690 Days
• Field of Search: 704 1, 704 9, 704 10
• CPC Class Codes: G06N 20/10 using kernel methods, e.g. support vector machines [SVM]
• CPC Class Codes: G06N 3/04 Architecture, e.g. interconnection topology
• CPC Class Codes: G06N 3/044 Recurrent networks, e.g. Hopfield networks
• CPC Class Codes: G06N 3/045 Combinations of networks
• CPC Class Codes: G06N 3/047 Probabilistic or stochastic networks
• CPC Class Codes: G06N 3/084 Backpropagation, e.g. using gradient descent
• CPC Class Codes: G06N 3/088 Non-supervised learning, e.g. competitive learning
• CPC Class Codes: G06Q 30/016 After-sales

Abstract

A company may desire to maintain a quality level for messages sent by customer service representatives to customers. The company may receive a message input by a customer service representative, modify the message with one or more neural networks, and transmit the modified message to a customer. To modify a message, an input vector may be created for each word of the message where the input vector is created using a word embedding of the word and a feature vector that represents the characters of the word. The input vectors for the words of the message may be sequentially processed with an encoding neural network to compute a message encoding vector that represents the message. The message encoding vector may then be processed by a decoding neural network to sequentially generate the words of a modified message. The modified message may then be transmitted to the customer.

Claims

1. A computer-implemented method for modifying a customer service message, the method comprising:

receiving words of an input message from a customer service representative, the input message comprising a message for a customer;

obtaining a word embedding vector for each of the words of the input message, wherein a word embedding vector represents a corresponding word of the input message in a vector space;

computing a character feature vector for each of the words of the input message, wherein a first character feature vector for a first word is computed by;

obtaining a representation of each character of the first word, and

sequentially processing the representations of the characters of the first word with a character neural network;

creating an input vector for each of the words of the input message, wherein creating a first input vector for the first word comprises combining;

a first word embedding vector for the first word, and

the first character feature vector for the first word;

computing a message encoding vector by sequentially processing the input vectors for the words of the input message with an encoding neural network;

generating a modified message with a decoding neural network, wherein generating the modified message comprises generating words of the modified message sequentially, and wherein the message encoding vector is an input to the decoding neural network; and

transmitting the modified message to the customer as a replacement for the input message.

2. The computer-implemented method of claim 1, wherein the character neural network comprises a convolutional layer.

3. The computer-implemented method of claim 1, wherein at least one of the encoding neural network or the decoding neural network comprises a recurrent neural network.

4. The computer-implemented method of claim 1, wherein at least one of the encoding neural network or the decoding neural network comprises a long short-term memory component.

5. The computer-implemented method of claim 1, wherein generating the modified message with the decoding neural network comprises using a beam search.

6. The computer-implemented method of claim 1, wherein generating the modified message with the decoding neural network comprises, for each word of the modified message:

computing a vocabulary distribution using a state vector of the decoding neural network; and

selecting one or more words using the vocabulary distribution.

7. The computer-implemented method of claim 1, wherein generating the modified message with the decoding neural network comprises, for each word of the modified message, computing an attention distribution over the words of the input message.

8. A system for modifying a customer service message, 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;

receive words of an input message provided by a customer service representative, the input message comprising a message for a customer;

obtain a word embedding vector for each of the words of the input message, wherein each word embedding vector represents a corresponding word in a vector space;

compute a character feature vector for each of the words of the input message, wherein computing a first character feature vector for a first word comprises obtaining representations of individual characters of the first word and processing the representations of the individual characters of the first word with a character neural network;

create an input vector for each of the words of the input message, wherein creating a first input vector for the first word comprises combining a first word embedding vector for the first word and the first character feature vector for the first word;

compute a message encoding vector by sequentially processing the input vectors for the words of the input message with an encoding neural network;

generate a modified message with a decoding neural network, wherein generating the modified message comprises generating words of the modified message sequentially, and wherein the message encoding vector is an input to the decoding neural network; and

cause the modified message to be transmitted to the customer as a replacement for the input message.

9. The system of claim 8, wherein:

the encoding neural network is configured to receive as input a first type vector that indicates that the input message is an unsuitable message; and

the decoding neural network is configured to receive as input a second type vector that indicates that the modified message is to be a suitable message.

10. The system of claim 9, wherein each of the encoding neural network and the decoding neural network comprise neural networks trained using an adversarial discriminator.

11. The system of claim 9, wherein the first type vector and the second type vector were computed during training of the encoding neural network and the decoding neural network.

12. The system of claim 8, wherein the at least one server computer is configured to generate the modified message with the decoding neural network by, for each word of the modified message, computing an attention distribution over the words of the input message.

13. The system of claim 12, wherein the at least one server computer is configured to generate the modified message with the decoding neural network by, for each word of the modified message, computing a probability for selecting a word from the input message.

14. The system of claim 8, wherein each of the encoding neural network and the decoding neural network comprise neural networks trained using an auto-encoder that modifies a suitable message to an unsuitable message and then modifies the unsuitable message back to another suitable message.

15. One or more non-transitory computer-readable media comprising computer executable instructions that, when executed, cause at least one processor to perform actions comprising:

receiving words of an input message entered by a customer service representative, the input message comprising a message for a customer;

obtaining a word embedding vector for each of the words of the input message, wherein each word embedding vector represents a corresponding word in a vector space;

computing a character feature vector for each of the words of the input message, wherein computing a first character feature vector for a first word comprises sequentially processing characters of the first word with a character neural network;

creating an input vector for each of the words of the input message, wherein creating a first input vector for the first word comprises combining a first word embedding vector for the first word and the first character feature vector for the first word;

computing a message encoding vector by sequentially processing the input vectors for the words of the input message with an encoding neural network;

generating a modified message with a decoding neural network, wherein generating the modified message comprises generating words of the modified message sequentially, and wherein the message encoding vector is an input to the decoding neural network; and

causing the modified message to be transmitted to the customer as a replacement for the input message.

16. The one or more non-transitory computer-readable media of claim 15, the actions comprising processing the input message with a message classifier to determine that the input message does not meet a quality level.

17. The one or more non-transitory computer-readable media of claim 16, wherein the message classifier processes the input vectors for the words of the input message.

18. The one or more non-transitory computer-readable media of claim 16, wherein the message classifier comprises a convolutional layer and a recurrent layer.

19. The one or more non-transitory computer-readable media of claim 15, wherein the first word is an out-of-vocabulary word and the first word embedding corresponds to a token for out-of-vocabulary words.

20. The one or more non-transitory computer-readable media of claim 15, wherein generating the modified message with the decoding neural network comprises, for each word of the modified message, computing an attention distribution over the words of the input message.

21. The one or more non-transitory computer-readable media of claim 15, wherein sequentially processing the characters of the first word with the character neural network comprises sequentially processing character embeddings or one-hot vectors of the characters of the first word.