Building and using target-based sentiment models

Patent metadata

Field	Value
Publication number	US-12321701-B2
Application number	US-202217981293-A
Country	US
Kind code	B2
Filing date	Nov 4, 2022
Priority date	Nov 4, 2022
Publication date	Jun 3, 2025
Grant date	Jun 3, 2025

Abstract

Systems and methods are directed to training and utilizing a generative language model that is constrained by a predetermined template that is used to train the generative language model. Once trained, customer data is accessed and transmitted to an evaluation component associated with the generative language model. The generative language model generates one or more sentences based on a feedback input of the plurality of feedback inputs, whereby the one or more sentences each include a sentiment, a target, and a reason for the sentiment in a format defined by the predetermined template. The evaluation component then identifies the sentiment, the target, and the reason from a sentence of the one or more sentences. A communication is then presented, on a device of a user, based on at least the sentiment and the reason identified from the sentence. The communication can be an alert or a report.

First claim

What is claimed is:

1. A method comprising:
training a generative language model to identify features that it was not explicitly trained with, the training the generative language model comprising:
accessing training data and annotated training data, the annotated training data comprising annotations of the training data constrained by a predetermined template to be in a specific format; and
training, by a training component, the generative language model to perform natural language generation tasks that include generating one or more sentences comprising a sentiment, a target, and a reason for the sentiment in a format defined by the predetermined template without having been trained on the target or reason;

accessing, by a communication interface of a network system, customer data that includes a plurality of feedback input;
transmitting the customer data to an evaluation component of the network system;
generating, by the generative language model associated with the evaluation component, the one or more sentences based on a feedback input of the plurality of feedback inputs, the one or more sentences each including the sentiment, the target, and the reason for the sentiment in the format defined by the predetermined template used to train the generative language model;
identifying, by the evaluation component, the sentiment, the target, and the reason from a sentence of the one or more sentences generated by the generative language model; and
automatically, without human intervention, surfacing, by a recommendation module, a recommendation for an action to be performed by a user based on the sentiment and the reason and on a table that indicates a plurality of outputs from the generative language model, a sentiment associated with each output, one or more tokens associated with a target or reason associated with each output, and a recommendation for an action to be performed for each output.

2. The method of claim 1 , wherein the format of the predetermined template is “The customer is <sentiment> about <target> because <reason>”.

3. The method of claim 1 , further comprising:
generating an alert that indicates a link to a ticket associated with the sentence, the sentiment, the target, and the reason; and
transmitting the alert to a device of the user.

4. The method of claim 1 , wherein the automatically surfacing the recommendation comprises transmitting the recommendation to a device of the user.

5. The method of claim 1 , further comprising:
generating the table that indicates the plurality of outputs from the generative model, the sentiment associated with each output, the one or more tokens associated with the target or reason associated with each output, and the recommendation to surface to a device of the user based on each output.

6. The method of claim 1 , wherein the training the generative language model comprises training a text-to-text-transfer-transformation (T5) model.

7. The method of claim 1 , wherein the training the generative language model comprises performing multi-task training for an email task and a survey task.

8. The method of claim 1 , wherein the training the generative language model further comprises randomly augmenting the training data or the annotated training data using a natural language processing (NPL) augmentation library.

9. The method of claim 8 , wherein the randomly augmenting comprises one or more of a contextual substitution, a contextual insertion, replacing a word with a synonym, or introducing a spelling mistake.

10. The method of claim 1 , wherein the training the generative language model comprises randomly choosing to ignore loss of tokens that do not belong to the sentiment or the target.

11. A system comprising:
one or more hardware processors; and
a memory storing instructions that, when executed by the one or more hardware processors, cause the one or more hardware processors to perform operations comprising:
training a generative language model to identify features that it was not explicitly trained with, the training the generative language model comprising:
accessing training data and annotated training data, the annotated training data comprising annotations of the training data constrained by a predetermined template to be in a specific format; and
training, by a training component, the generative language model to perform natural language generation tasks that include generating one or more sentences comprising a sentiment, a target, and a reason for the sentiment in a format defined by the predetermined template without having been trained on the target or reason;

accessing, by a communication interface of a network system, customer data that includes a plurality of feedback input;
transmitting the customer data to an evaluation component of the network system;
generating, by the generative language model associated with the evaluation component, the one or more sentences based on a feedback input of the plurality of feedback inputs, the one or more sentences each including the sentiment, the target, and the reason for the sentiment in the format defined by the predetermined template used to train the generative language model;
identifying, by the evaluation component, the sentiment, the target, and the reason from a sentence of the one or more sentences generated by the generative language model; and

automatically, without human intervention, surfacing, by a recommendation module, a recommendation for an action to be performed by a user based on the sentiment and the reason and on a table that indicates a plurality of outputs from the generative language model, a sentiment associated with each output, one or more tokens associated with a target or reason associated with each output, and a recommendation for an action to be performed for each output.

12. The system of claim 11 , wherein the format of the predetermined template is “<sentiment> about <target> because <reason>”.

13. The system of claim 11 , wherein the operations further comprise:
generating an alert that indicates a link to a ticket associated with the sentence, the sentiment, the target, and the reason; and
transmitting the alert to a device of the user.

14. The system of claim 11 , wherein the automatically surfacing the recommendation comprises
transmitting the recommendation to a device of the user.

15. He system of claim 11 , wherein the training the generative language model comprises performing multi-task training for an email task and a survey task.

16. The system of claim 11 , wherein the training the generative language model further comprises randomly augmenting the training data or the annotated training data prior to training the generative language model.

17. The system of claim 11 , wherein the training the generative language model comprises randomly choosing to ignore loss of tokens that do not belong to the sentiment or the target.

18. A machine-storage medium comprising instructions which, when executed by one or more hardware processors of a machine, cause the machine to perform operations comprising:
training a generative language model to identify features that it was not explicitly trained with, the training the generative language model comprising:
accessing training data and annotated training data, the annotated training data comprising annotations of the training data constrained by a predetermined template to be in a specific format; and
training, by a training component, the generative language model to perform natural language generation tasks that include generating

Assignees

• Microsoft Technology Licensing Llc

Inventors

• Anand Vishal
• Mishra Ananya
• Ballapuram Pramodith
• Wu Cheng

Classifications

• G06Q30/0282

  Rating or review of business operators or products · CPC title

• G06Q30/0203

  Market surveys; Market polls · CPC title

• G06F40/56

  Natural language generation · CPC title

• G06Q30/0202

  Market predictions or forecasting for commercial activities · CPC title

• G06Q10/067

  Enterprise or organisation modelling · CPC title