Automated support sub-topic classification using large language models

1 . A method for automated support sub-topic classification, comprising:

collecting, by at least one processor, user transcripts and large language model (LLM) generated summaries of the transcripts;

performing, by the at least one processor, unsupervised learning to identify common themes in both the transcripts and the LLM generated summaries;

generating, by the at least one processor, sub-topics by analyzing the identified common themes through the unsupervised learning;

creating, by the at least one processor, labeled datasets by pairing the transcripts with the LLM generated summaries and the sub-topics generated from the common themes;

training, by the at least one processor, a supervised learning model using the labeled datasets to perform categorization of the user transcripts into the sub-topics;

performing, by the at least one processor using the trained supervised learning model, real-time categorization of a new user query into the sub-topics by generating an LLM summary of the new user query and inputting both the new user query and the LLM summary of the new user query into the trained supervised learning model; and

generating, by the at least one processor, an action response to the new user query by selecting the action response corresponding to the categorization.

2 . The method of claim 1 , wherein preprocessing the collected transcripts comprises:

normalizing text in the user transcripts; and

tokenizing, by the LLM, the user transcripts prior to the LLM generating the summaries of the transcripts.

3 . The method of claim 1 , wherein performing the unsupervised learning comprises:

applying a clustering algorithm to the transcripts and the summaries to identify groups of similar content.

4 . The method of claim 1 , wherein generating the sub-topics comprises:

analyzing the identified common themes;

grouping related themes; and

assigning descriptive labels to each group of related themes.

5 . The method of claim 1 , wherein training the supervised learning model comprises:

splitting the labeled datasets into training and validation sets;

applying a machine learning algorithm to the training set;

evaluating the model's performance using the validation set; and

periodically retraining the supervised learning model with new labeled datasets to improve categorization accuracy.

6 . The method of claim 1 , wherein generating the action response includes one or more of:

providing an automated answer to the query;

routing the query to an appropriate support agent;

escalating the query to a higher tier of support;

requesting additional information from the user;

providing relevant documentation or resources;

initiating a troubleshooting process;

scheduling a follow-up interaction; and

updating a knowledge base with information from the query and response.

7 . The method of claim 1 , wherein performing the real-time categorization comprises:

extracting features from the new user query;

applying the trained model to the extracted features; and

assigning a sub-topic category to the new user query by selecting the sub-topic category output by the model.

8 . The method of claim 7 , wherein generating the action response to the new user query comprises:

retrieving pre-defined action responses associated with the assigned sub-topic category;

selecting an appropriate action response by analyzing content of the new user query; and

customizing the selected action response to address the new user query.

9 . The method of claim 1 , further comprising:

analyzing user feedback on the generated action responses; and

updating the supervised learning model by incorporating the feedback into the training.

10 . The method of claim 1 , further comprising:

generating performance metrics for the categorization;

identifying the sub-topics with a categorization accuracy below a predetermined threshold according to the performance metrics; and

refining the sub-topics and the labeled datasets for the sub-topics with the categorization accuracy below the predetermined threshold.

11 . A system for automated support sub-topic classification, comprising:

a processor; and

a memory storing instructions that, when executed by the processor, cause the system to:

collect, by a data collection module, user transcripts and large language model (LLM) generated summaries of the transcripts;

perform, by an unsupervised learning module, unsupervised learning to identify common themes in both the transcripts and the LLM generated summaries;

generate, by a sub-topic generation module, sub-topics by analyzing the identified common themes through the unsupervised learning;

create, by a dataset creation module, labeled datasets by pairing the transcripts with the LLM generated summaries and the sub-topics generated from the common themes;

train, by a model training module, a supervised learning model using the labeled datasets to perform categorization of the user transcripts into the sub-topics;

perform, by the trained supervised learning model, real-time categorization of a new user query into the sub-topics by generating an LLM summary of the new user query and inputting both the new user query and the LLM summary of the new user query into the trained supervised learning model; and

generate, by an action response module, an action response to the new user query by selecting the action response corresponding to the categorization.

12 . The system of claim 11 , wherein preprocessing the collected transcripts comprises:

normalizing, by a text normalization module, text in the user transcripts; and

tokenizing, by the LLM, the user transcripts prior to the LLM generating the summaries of the transcripts.

13 . The system of claim 11 , wherein performing the unsupervised learning comprises applying, by the unsupervised learning module, a clustering algorithm to the transcripts and the summaries to identify groups of similar content.

14 . The system of claim 11 , wherein generating the sub-topics comprises:

analyzing, by the sub-topic generation module, the identified common themes;

grouping, by the sub-topic generation module, related themes; and

assigning, by the sub-topic generation module, descriptive labels to each group of related themes.

15 . The system of claim 11 , wherein training the supervised learning model comprises:

splitting, by the model training module, the labeled datasets into training and validation sets;

applying, by the model training module, a machine learning algorithm to the training set;

evaluating, by the model training module, the model's performance using the validation set; and

periodically retraining, by the model training module, the supervised learning model with new labeled datasets to improve categorization accuracy.

16 . The system of claim 11 , wherein generating the action response includes one or more of:

providing, by the action response module, an automated answer to the query;

routing, by the action response module, the query to an appropriate support agent;

escalating, by the action response module, the query to a higher tier of support;

requesting, by the action response module, additional information from the user;

providing, by the action response module, relevant documentation or resources;

initiating, by the action response module, a troubleshooting process;

scheduling, by the action response module, a follow-up interaction; and

updating, by the action response module, a knowledge base with information from the query and response.

17 . The system of claim 11 , wherein performing the real-time categorization comprises:

extracting, by a feature extraction module, features from the new user query;

applying, by the trained model, the trained model to the extracted features; and

assigning, by the trained model, a sub-topic category to the new user query by selecting the sub-topic category output by the model.

18 . The system of claim 17 , wherein generating the action response to the new user query comprises:

retrieving, by the action response module, pre-defined action responses associated with the assigned sub-topic category;

selecting, by the action response module, an appropriate action response by analyzing content of the new user query; and

customizing, by the action response module, the selected action response to address the new user query.

19 . The system of claim 11 , wherein the instructions further cause the system to:

analyze, by a feedback analysis module, user feedback on the generated action responses; and

update, by the model training module, the supervised learning model by incorporating the feedback into the training.

20 . A method for automated support sub-topic classification, comprising:

receiving, by at least one processor, a new user query;

performing, by the at least one processor using a trained supervised learning model, real-time categorization of the new user query into sub-topics by generating a large language model (LLM) summary of the new user query and inputting both the new user query and the LLM summary of the new user query into the trained supervised learning model, wherein the trained supervised learning model is trained using labeled datasets created by pairing user transcripts with the LLM generated summaries and the sub-topics generated based on common themes identified through unsupervised learning of both the user transcripts and the LLM generated summaries; and

generating, by the at least one processor, an action response to the new user query in response to the categorization.