Granted Patent B2

US 12,651,284 · App. 17/968,714 · Granted Jun 9, 2026

Inventors: David Chen (Durham, NC); Maury Courtland (Durham, NC); Aysu Ezen Can (Cary, NC); Sahil Badyal (McLean, VA)

Assignee: CAPITAL ONE SERVICES, LLC

G06Q30/0282G06F3/0484G06F40/30G06N20/20

Abstract

Claims (74)

1 . A customer satisfaction detection system comprising:

one or more processors;

memory in communication with the one or more processors and storing instructions that are configured to cause the customer satisfaction detection system to:

receive aspect sentiment training data, wherein the aspect sentiment training data has a training polarity of positive, negative or neutral;

train a machine learning model using the aspect sentiment training data by:

randomly selecting a token in a tokenized original sequence of the training data;

converting the selected token into a mask; and

unmasking the token using the machine learning model;

receive call data;

extract one or more aspects from the call data;

apply the machine learning model to the one or more aspects from the call data, wherein the machine learning model outputs a table of the one or more aspects with a sentiment polarity of positive, negative, or neutral;

filter the one or more aspects and the associated sentiment polarity into categories;

analyze for trends based on the filtered aspects and sentiment polarity;

generate a dynamic graphical user interface for displaying the trends; and

send the dynamic graphical user interface to a user device for display.

2 . The customer satisfaction detection system of claim 1 , wherein the categories are determined using a top-down approach.

3 . The customer satisfaction detection system of claim 2 , wherein the categories are determined using hierarchical embedding.

4 . The customer satisfaction detection system of claim 1 , wherein the categories are determined using a bottom-up approach.

5 . The customer satisfaction detection system of claim 1 , wherein the machine learning model is a transformer-based neural network.

6 . The customer satisfaction detection system of claim 1 , wherein the machine learning model is a pretrained bidirectional encoder representation from transformers (BERT) model.

7 . The customer satisfaction detection system of claim 6 , wherein training the BERT model further includes using an unsupervised data augmentation (UDA) module.

8 . The customer satisfaction detection system of claim 1 , wherein filtering the aspects and sentiment polarity into categories comprises a pair of words with variations, wherein the customer satisfaction detection system is configured to score the pair of words the same despite the variations using the machine learning model.

9 . The customer satisfaction detection system of claim 1 , wherein the machine learning model processes the aspect sentiment training data at a span level.

10 . The customer satisfaction detection system of claim 1 , wherein the aspects are a noun, a noun-phrase, or both, and wherein the memory stores further instructions that are configured to cause the customer satisfaction detection system to:

convert audio data to text data.

11 . A generic aspect-based sentiment analysis system comprising:

one or more processors;

memory in communication with the one or more processors and storing instructions that are configured to cause the generic aspect-based sentiment analysis system to:

receive training data, wherein the training data has a training polarity of positive, negative or neutral;

train a pretrained bidirectional encoder representation from transformers (BERT) model with an unsupervised data augmentation (UDA) module at a span level with the training data by:

randomly selecting a token in a tokenized original sequence of the training data;

converting the selected token into a mask; and

unmasking the token using the BERT model;

receive data;

extract aspects from the data;

apply the BERT model to one or more aspects from the data, wherein the BERT model outputs a table of the one or more aspects with a sentiment polarity of positive, negative, or neutral;

filter the aspects and the sentiment polarity into categories;

analyze trends based on the filtered aspects and sentiment polarity; and

output a result based on the trends.

12 . The generic aspect-based sentiment analysis system of claim 11 , wherein the UDA processes the training data at the span level further comprises:

verifying that unmasked token does not have the same value as the token prior to being unmasked; and

verifying that the unmasked token is excluded from a set containing punctuation.

13 . The generic aspect-based sentiment analysis system of claim 11 , wherein randomly selecting the token in the tokenized original sequence of the training data further comprises avoiding punctuation of the training data.

14 . The generic aspect-based sentiment analysis system of claim 11 , wherein training the BERT model is iterated a number of times related to a length of the tokenized original sequence.

15 . The generic aspect-based sentiment analysis system of claim 14 , further comprising:

determining whether the unmasked tokens that have values greater than a confidence threshold; and

further training the BERT model with the unmasked tokens that have values greater than the confidence threshold.

16 . A generic aspect-based sentiment analysis system comprising:

one or more processors;

memory in communication with the one or more processors and storing instructions that are configured to cause the generic aspect-based sentiment analysis system to:

receive training data, wherein the training data has a training polarity of positive, negative or neutral;

train a machine learning model with the training data at a span level by:

randomly selecting a token in a tokenized original sequence of the training data;

converting the selected token into a mask; and

unmasking the token using the machine learning model;

receive data;

extract aspects from data;

apply the machine learning model to one or more aspects from the data, wherein the machine learning model outputs a table of the one or more aspects with a sentiment polarity of positive, negative, or neutral;

filter the aspects and the sentiment polarity into categories;

analyze for trends based on the filtered aspects and sentiment polarity; and

output a result based on the trends.

17 . The generic aspect-based sentiment analysis system of claim 16 , wherein the machine learning model is a pretrained bidirectional encoder representation from transformers (BERT) model, and training the BERT model further includes using an unsupervised data augmentation (UDA) module that further processes training data at the span level by:

verifying that an unmasked token does not have the same value as the token prior to being unmasked; and

verifying that the unmasked token is excluded from a set containing punctuation.

18 . The generic aspect-based sentiment analysis system of claim 17 , wherein:

training the BERT model is iterated a number of times related to a length of the tokenized original sequence;

determining whether the unmasked tokens that have values greater than a confidence threshold; and

further training the BERT model with the unmasked tokens that have values greater than the confidence threshold.

19 . The generic aspect-based sentiment analysis system of claim 16 , wherein the memory stores further instructions that are configured to cause the generic aspect-based sentiment analysis system to:

generate a dynamic graphical user interface displaying the trends;

send the dynamic graphical user interface to a user device for display;

receive, from the user device, modifications to the trends;

generate a modified dynamic graphical user interface displaying modified trends; and

send the modified dynamic graphical user interface to the user device for display.

Assignments (1)

Continuity (2)

References Cited (35)