Intent discovery using language models
A method for providing an intent expressed in a conversational interaction in narrative form includes detecting a first intent, with a first large language model, from an input comprising at least an interaction transcript and an engineered prompt; generating, with the first large language model, a first narrative for the first intent and a first confidence score corresponding to a probability that the first intent is an intention in fact expressed in the interaction transcript; determining that the first confidence score is greater than or equal to a threshold; checking that the first narrative conforms to fluency and completeness rules, when the first confidence score is greater than or equal to the threshold; and outputting the first narrative as an output detected intent, when the first confidence score is greater than or equal to the threshold and when the first narrative conforms to the fluency and completeness rules.
1 . A method for providing an intent expressed in a conversational interaction in narrative form, comprising:
detecting a first intent, with a first large language model, from an input comprising at least an interaction transcript and an engineered prompt;
generating, with the first large language model, a first narrative for the first intent and a first confidence score corresponding to a probability that the first intent is an intention in fact expressed in the interaction transcript;
determining that the first confidence score is greater than or equal to a threshold;
checking that the first narrative for the first intent conforms to one or more fluency and completeness rules, when the first confidence score is determined to be greater than or equal to the threshold; and
outputting the first narrative as an output detected intent, when the first confidence score is determined to be greater than or equal to the threshold and when the first narrative for the first intent conforms to the one or more fluency and completeness rules.
2 . The method of claim 1 , further comprising:
modifying one or more hyperparameters of the first large language model when the first narrative for the first intent does not conform to the one or more fluency and completeness rules;
causing the first large language model to implement the modified one or more hyperparameters and to generate an updated first narrative for the first intent;
checking that the updated first narrative conforms to the one or more fluency and completeness rules; and
outputting the updated first narrative as the output detected intent, when the updated first narrative conforms to the one or more fluency and completeness rules.
3 . The method of claim 1 , further comprising:
determining that the first confidence score is less than the threshold;
detecting a second intent, with a second large language model, from the input;
generating, with the second large language model, a second narrative for the second intent and a second confidence score corresponding to a probability that the second intent is an intention in fact expressed in the interaction transcript;
checking that the second narrative for the second intent conforms to the one or more fluency and completeness rules; and
outputting the second narrative as the output detected intent, when the second narrative for the second intent conforms to the one or more fluency and completeness rules.
4 . The method of claim 3 , further comprising:
modifying one or more hyperparameters of the second large language model when the second narrative for the second intent does not conform to the one or more fluency and completeness rules;
causing the second large language model to implement the modified one or more hyperparameters and to generate an updated second narrative for the second intent;
checking that the updated second narrative conforms to the one or more fluency and completeness rules; and
outputting the updated second narrative as the output detected intent, when the updated second narrative conforms to the one or more fluency and completeness rules.
5 . The method of claim 1 , further comprising:
selecting the first narrative for evaluation by a third large language model;
evaluating the input corresponding to the first narrative to generate a validated detected intent in narrative form; and
outputting the validated detected intent in narrative form as the output detected intent.
6 . The method of claim 1 , further comprising trimming a length of the first narrative to be equal to or less than a predefined length.
7 . The method of claim 1 , wherein the one or more fluency and completeness rules comprises at least one of a grammar rule, a spelling rule, or a punctuation rule.
8 . The method of claim 1 , wherein the engineered prompt is automatically or manually selected from a predefined list of engineered prompts.
9 . An apparatus configured for providing an intent expressed in a conversational interaction in narrative form, comprising: one or more memories comprising processor-executable instructions; and one or more processors configured to execute the processor-executable instructions and cause the apparatus to:
detect a first intent, with a first large language model, from an input comprising at least an interaction transcript and an engineered prompt;
generate, with the first large language model, a first narrative for the first intent and a first confidence score corresponding to a probability that the first intent is an intention in fact expressed in the interaction transcript;
determine that the first confidence score is greater than or equal to a threshold;
check that the first narrative for the first intent conforms to one or more fluency and completeness rules, when the first confidence score is determined to be greater than or equal to the threshold; and
output the first narrative as an output detected intent, when the first confidence score is determined to be greater than or equal to the threshold and when the first narrative for the first intent conforms to the one or more fluency and completeness rules.
10 . The apparatus of claim 9 , wherein the one or more processors are configured to execute the processor-executable instructions and cause the apparatus to:
modify one or more hyperparameters of the first large language model when the first narrative for the first intent does not conform to the one or more fluency and completeness rules;
cause the first large language model to implement the modified one or more hyperparameters and to generate an updated first narrative for the first intent;
check that the updated first narrative conforms to the one or more fluency and completeness rules; and
output the updated first narrative as the output detected intent, when the updated first narrative conforms to the one or more fluency and completeness rules.
11 . The apparatus of claim 9 , wherein the one or more processors are configured to execute the processor-executable instructions and cause the apparatus to:
determine that the first confidence score is less than the threshold;
detect a second intent, with a second large language model, from the input;
generate, with the second large language model, a second narrative for the second intent and a second confidence score corresponding to a probability that the second intent is an intention in fact expressed in the interaction transcript;
check that the second narrative for the second intent conforms to the one or more fluency and completeness rules; and
output the second narrative as the output detected intent, when the second narrative for the second intent conforms to the one or more fluency and completeness rules.
12 . The apparatus of claim 11 , wherein the one or more processors are configured to execute the processor-executable instructions and cause the apparatus to:
modify one or more hyperparameters of the second large language model when the second narrative for the second intent does not conform to the one or more fluency and completeness rules;
cause the second large language model to implement the modified one or more hyperparameters and to generate an updated second narrative for the second intent;
check that the updated second narrative conforms to the one or more fluency and completeness rules; and
output the updated second narrative as the output detected intent, when the updated second narrative conforms to the one or more fluency and completeness rules.
13 . The apparatus of claim 9 , wherein the one or more processors are configured to execute the processor-executable instructions and cause the apparatus to:
select the first narrative for evaluation by a third large language model;
evaluate the input corresponding to the first narrative to generate a validated detected intent in narrative form; and
output the validated detected intent in narrative form as the output detected intent.
14 . The apparatus of claim 9 , wherein the one or more processors are configured to execute the processor-executable instructions and cause the apparatus to trim a length of the first narrative to be equal to or less than a predefined length.
15 . A method for categorization of intents, comprising:
receiving, with an embedding component, a plurality of intents corresponding to conversational interactions;
converting, with the embedding component, textual data structure of the plurality of intents into numerical vectors embeddings;
discerning, with a categorization component processing the numerical vectors embeddings of the plurality of intents, one or more categories that are present within the plurality of intents; and
labeling, with a label generation component, the one or more categories with a keyphrase.
16 . The method of claim 15 , wherein the keyphrase generated by the label generation component is a phrase extracted from an overlapping portion of the plurality of intents categorized within each of the one or more categories.
17 . The method of claim 15 , wherein the keyphrase generated by the label generation component is an abstraction based on the plurality of intents categorized within each of the one or more categories.
18 . The method of claim 15 , further comprising detecting at least one intent of the plurality of intents with a first large language model from an input comprising at least an interaction transcript and an engineered prompt.
19 . The method of claim 15 , further comprising generating, with a visualization component, a visualization of the one or more categories.
20 . The method of claim 19 , wherein the visualization is a treemap.