Inverse text normalization of contact center communications

1. A method for inverse text normalization of contact center communications, the method comprising:

performing named entity recognition on text from a contact center communication to identify one or more entities in the text;

normalizing each of the identified one or more entities in the text using weighted finite-state transducers; and

normalizing at least one entity identified in the text using a large language model in response to determining that the at least one entity identified in the text was unable to be normalized using the weighted finite-state transducers; wherein normalizing the at least one entity identified in the text using the large language model comprises:

calculating an average of probabilities from word-probability pairs output by the large language model;

determining whether the average exceeds a predefined confidence threshold; and

accepting a normalization output of the large language model in response to determining that the average exceeds the predefined confidence threshold.

2. The method of claim 1 , wherein the contact center communication comprises a speech-to-text transcript of at least a portion of a contact center conversation with a user.

3. The method of claim 1 , wherein normalizing each of the identified one or more entities in the text using weighted finite-state transducers comprises normalizing a plurality of identified entities in the text using weighted finite-state transducers.

4. The method of claim 3 , wherein normalizing the plurality of identified entities in the text using weighted finite-state transducers comprises using a dedicated graph for each entity type of the plurality of identified entities.

5. The method of claim 1 , wherein normalizing the at least one entity identified in the text using the large language model comprises:

receiving, for each entity of the at least one entity, entity text associated with the respective entity identified in the text;

comparing the entity text to a set of training examples for the large language model;

selecting training examples from the set of training examples that are most similar to the entity text; and

applying the large language model to the entity text using a large language model prompt that includes the selected training examples.

6. The method of claim 5 , wherein selecting the training examples from the set of training examples that are most similar to the entity text comprises selecting training examples from the set of training examples that share the most common words.

7. The method of claim 6 , wherein selecting the training examples from the set of training examples that are most similar to the entity text comprises selecting at least five training examples.

8. The method of claim 1 , further comprising rejecting the normalization output of the large language model in response to determining that the average does not exceed the predefined confidence threshold.

9. The method of claim 1 , wherein performing the named entity recognition on the text from the contact center communication to identify the one or more entities in the text comprises identifying at least one of a date, a phone number, a currency, a social security number, a time of day, or a zip code.

10. A computing system for inverse text normalization of contact center communications, the computing system comprising:

at least one processor; and

at least one memory comprising a plurality of instructions stored thereon that, in response to execution by the at least one processor, causes the computing system to:

perform named entity recognition on text from a contact center communication to identify one or more entities in the text;

normalize each of the identified one or more entities in the text using weighted finite-state transducers; and

normalize at least one entity identified in the text using a large language model in response to a determination that the at least one entity identified in the text was unable to be normalized using the weighted finite-state transducers; wherein to normalize the at least one entity identified in the text using the large language model comprises to:

calculate an average of probabilities from word-probability pairs output by the large language model;

determine whether the average exceeds a predefined confidence threshold;

accept a normalization output of the large language model in response to a determination that the average exceeds the predefined confidence threshold; and

reject the normalization output of the large language model in response to a determination that the average does not exceed the predefined confidence threshold.

11. The computing system of claim 10 , wherein the contact center communication comprises a speech-to-text transcript of at least a portion of a contact center conversation with a user.

12. The computing system of claim 10 , wherein to normalize each of the identified one or more entities in the text using weighted finite-state transducers comprises to normalize a plurality of identified entities in the text using weighted finite-state transducers.

13. The computing system of claim 12 , wherein to normalize the plurality of identified entities in the text using weighted finite-state transducers comprises to use a dedicated graph for each entity type of the plurality of identified entities; and

wherein the plurality of identified entities represents a plurality of different entity types.

14. The computing system of claim 10 , wherein to normalize the at least one entity identified in the text using the large language model comprises to:

receive, for each entity of the at least one entity, entity text associated with the respective entity identified in the text;

compare the entity text to a set of training examples for the large language model;

select training examples from the set of training examples that are most similar to the entity text; and

apply the large language model to the entity text using a large language model prompt that includes the selected training examples.

15. The computing system of claim 14 , wherein to select the training examples from the set of training examples that are most similar to the entity text comprises to select training examples from the set of training examples that share the most common words.

16. One or more non-transitory machine-readable storage media comprising a plurality of instructions stored thereon that, in response to execution by a processor, causes a computing system to:

perform named entity recognition on text from a contact center communication to identify a plurality of entities in the text;

normalize each entity of the plurality of entities identified in the text using weighted finite-state transducers and using a dedicated graph for each entity type of the plurality of entities identified in the text;

determine that at least one entity identified in the text was unable to be normalized using the weighted finite-state transducers; and

normalize the at least one entity identified in the text using a large language model; wherein to normalize the at least one entity identified in the text using the large language model comprises to:

calculate an average of probabilities from word-probability pairs output by the large language model;

determine whether the average exceeds a predefined confidence threshold;

accept a normalization output of the large language model in response to a determination that the average exceeds the predefined confidence threshold; and

reject the normalization output of the large language model in response to a determination that the average does not exceed the predefined confidence threshold.

17. The one or more non-transitory machine-readable storage media of claim 16 , wherein to normalize the at least one entity identified in the text using the large language model comprises to:

receive, for each entity of the at least one entity, entity text associated with the respective entity identified in the text;

compare the entity text to a set of training examples for the large language model;

select training examples from the set of training examples that are most similar to the entity text; and

apply the large language model to the entity text using a large language model prompt that includes the selected training examples.