Automated call summarization based on filtered utterances

Patent metadata

Field	Value
Publication number	US-12475884-B2
Application number	US-202318312201-A
Country	US
Kind code	B2
Filing date	May 4, 2023
Priority date	May 4, 2023
Publication date	Nov 18, 2025
Grant date	Nov 18, 2025

Abstract

Systems and methods for automated summarization of customer service calls in a specialized field, such as medical services and insurance, are disclosed. Call transcripts are processed to determine the utterances and corresponding speakers in a customer service call. A feature vector is generated for each utterance and processed through a trained intent model to assign intent labels. The utterances are then filtered based on the intent labels to generate an extractive summary. The extractive summary is processed through a trained summarization model to generate a natural language summary for the customer service call.

First claim

What is claimed is:

1 . A computer-implemented method, comprising:
determining a training data pool comprised of a plurality of call transcript data objects; determining a plurality of training utterances and corresponding feature vectors for the plurality of call transcript data objects; iteratively determining a set of weak supervision intent labelling rules for training a trained intent model; iteratively processing the plurality of training utterances and corresponding feature vectors through the trained intent model to determine weighting coefficients for the trained intent model and at least one model acceptance metric; generating a master set of intent labels for the trained intent model; accessing a call transcript data object for a customer service call; determining a plurality of utterances attributed to a plurality of speakers in the customer service call; generating a feature vector for each utterance of the plurality of utterances; processing each feature vector through the trained intent model to assign an intent label to each utterance; filtering, based on a filter set of intent labels, the plurality of utterances to generate an extractive summary data object, wherein:
the filter set of intent labels is a selected subset of the master set of intent labels; and
the extractive summary data object comprises;
a subset of the plurality of utterances; and
corresponding intent labels and speaker identifiers for the subset of the plurality of utterances; and

generating, based on the extractive summary data object and a trained summarization model, a natural language summary data object for the customer service call.

2 . The computer-implemented method of claim 1 , further comprising:
associating the natural language summary data object with a call record for the customer service call; and automatically displaying, on an interface display of a customer service representative computing system, a visual representation of the natural language summary data object in response to the customer service representative computing system accessing the call record.

3 . The computer-implemented method of claim 1 , further comprising:
classifying, based on the extractive summary data object and a trained classification model, a call topic of the customer service call; and associating a call topic label for the call topic with the natural language summary data object.

4 . The computer-implemented method of claim 1 , further comprising:
accessing an audio data object for the customer service call; determining, from the audio data object and using a trained speaker segmentation model, the plurality of speakers in the customer service call; converting, using a trained speech-to-text model, the audio data object to the call transcript data object; labelling the plurality of speakers in the call transcript data object; and embedding the plurality of utterances in the call transcript data object, wherein determining the plurality of utterances is based on embeddings in the call transcript data object.

5 . The computer-implemented method of claim 1 , further comprising:
generating, from the natural language summary data object, an embedding and at least one corresponding feature vector for the natural language summary data object; and storing, associated with a member identifier for the customer service call, the embedding and at least one corresponding feature vector for processing through at least one automated processing model.

6 . The computer-implemented method of claim 5 , further comprising:
determining a plurality of claim identifiers associated with the member identifier; determining, based on the natural language summary data object and a plurality of claim records corresponding to the plurality of claim identifiers, at least one claim correlation value for the natural language summary data object; determining, based on a comparison of the at least one claim correlation value and a claim relatedness threshold, at least one related claim identifier from the plurality of claim identifiers; and associating the at least one related claim identifier with a call record for the customer service call.

7 . The computer-implemented method of claim 5 , further comprising:
determining at least one prior call record associated with the member identifier and comprising at least one prior natural language summary data object; determining, based on the natural language summary data object and the at least one prior natural language summary data object, at least one call correlation value for the natural language summary data object; determining, based on a comparison of the at least one call correlation value and a call relatedness threshold, at least one related call record from the at least one prior call record; associating the at least one related call record with a call record for the customer service call; and automatically displaying, on an interface display of a customer service representative computing system, a visual representation of the natural language summary data object in response to the customer service representative computing system accessing the at least on related call record.

8 . The computer-implemented method of claim 5 , further comprising:
extracting, from the natural language summary data object, at least one key term having a key term type; determining, based on the member identifier and the key term type, a plurality of data records comprising corresponding key terms; determining variance values among the at least one key term and the corresponding key terms; determining, based on the variance values, the at least one key term is in error; and correcting, based on the corresponding key terms, the at least on key term in the natural language summary data object.

9 . The computer-implemented method of claim 5 , further comprising:
associating the embedding and at least one corresponding feature vector for the natural language summary data object with a member feature set; processing, using the at least one automated processing model, the member feature set to determine a next best action for the member identifier; and automatically initiating, based on the next best action, communication to a member associated with the member identifier.

10 . A system, comprising:
one or more processors; and a memory storing instructions that, when executed, cause the one or more processors to:
access a call transcript data object for a customer service call;
determine a plurality of utterances attributed to a plurality of speakers in the customer service call;
generate a feature vector for each utterance of the plurality of utterances;
process each feature vector through a trained intent model to assign an intent label to each utterance;
filter, based on a filter set of intent labels, the plurality of utterances to generate an extractive summary data object, wherein the extractive summary data object comprises:
a subset of the plurality of utterances; and
corresponding intent labels and speaker identifiers for the subset of the plurality of utterances;

generate, based on the extractive summary data object and a trained summarization model, a natural language summary data object for the customer service call;
generate, from the natural language summary data object, an embedding and at least one corresponding feature vector for the natural language summary data object;
store, associated with a member identifier for the customer service call, the embedding and at least one corresponding feature vector for processing through at least one automated processing model;
associate the embedding and at least one corresponding featur

Assignees

• Aetna Inc

Inventors

• Churgin Matthew A
• Hasan Sheikh Sadid Al
• Telukuntla Prasad

Classifications

• H04M3/5175 Primary

  Call or contact centers supervision arrangements · CPC title

• G10L17/18

  Artificial neural networks; Connectionist approaches · CPC title

• G10L15/16

  using artificial neural networks · CPC title

• G10L15/26

  Speech to text systems (G10L15/08 takes precedence) · CPC title

• G10L15/02

  Feature extraction for speech recognition; Selection of recognition unit · CPC title