Classifying feedback from transcripts

A system comprising:

• at least one processor;
• an application, which when executed by the at least one processor:
• trains a language model to output a predicted word of an utterance, using training utterances from a set of training transcripts, wherein training the language model comprises:
• converting, by an encoder layer of the language model, the training utterances to encoder values,
• generating, by an attention layer of the language model, attention values using the encoder values, and
• generating the predicted word of the utterance for the training utterances based on the attention values, to obtain a trained language model;
• selects a set of selected transcripts from the set of training transcripts based on training transcript scores;
• filters the training utterances of the set of selected transcripts using the attention values obtained from the trained language model to obtain filtered training utterances;
• clusters the filtered training utterances using the encoder values obtained from the trained language model to obtain clusters of training utterances;
• groups the clusters of training utterances into topics;
• assigns topic labels to the topics to generate topic label-training utterance pairs;
• trains a classifier model with the topic label-training utterance pairs to output a predicted topic label for a training utterance;
• receives an utterance from a transcript from a communication session;
• processes the utterance with the classifier model to identify a topic label for the utterance;
• and
• routes the communication session using the topic label for the utterance.

A method comprising:

• processing training transcript scores to select a set of selected transcripts from a set of training transcripts;
• training a language model with next word prediction and masking using training utterances from a set of training transcripts to obtain a trained language model;
• processing the set of selected transcripts using attention values of the trained language model to filter the training utterances, wherein the attention values are obtained from an attention layer of the trained language model;
• clustering encoder values of the training utterances to generate clusters of the training utterances, wherein the encoder values are obtained from an encoding layer of the trained language model and wherein the attention layer is distinct from the encoding layer;
• assigning topic labels to topics, wherein one of the topics corresponds to one or more of the clusters;
• and
• training a classifier model to identify topic labels for the training utterances by using a processor to update one or more weights of the classifier model, wherein the training utterances are filtered using the attention values for the training utterances from the attention layer of the trained language model and are clustered using the encoder values for the training utterances from the encoder layer of the trained language model.