dialoguekit.nlg.template_from_training_data

Extract user response utterance templates from annotated training data.

Functions

build_template_from_instances(→ Dict[str, ...)

Builds the NLG template.

extract_utterance_template(→ Dict[str, ...)

Extracts utterance templates for each intent.

Module Contents

dialoguekit.nlg.template_from_training_data.build_template_from_instances(utterances: List[dialoguekit.core.annotated_utterance.AnnotatedUtterance]) → Dict[str, List[dialoguekit.core.annotated_utterance.AnnotatedUtterance]]

Builds the NLG template.

The dialogue acts from the utterances will be used to build the templates. If an utterance has multiple dialogue acts, the intents will be concatenated with a semicolon. If an utterance does not have any dialogue acts, it will be skipped.

Parameters:

utterances – List of utterances to build the template from.

Returns:

Dictionary with intents as keys and lists of corresponding templates.

dialoguekit.nlg.template_from_training_data.extract_utterance_template(annotated_dialogue_file: str, participant_to_learn: str = 'USER', satisfaction_classifier: None | dialoguekit.nlu.models.satisfaction_classifier.SatisfactionClassifier | None = None) → Dict[str, List[dialoguekit.core.annotated_utterance.AnnotatedUtterance]]

Extracts utterance templates for each intent.

If a Satisfaction Classifier is provided it will be used to classify the utterances. The classification logic is as follows:

  • Hold participant utterance.

  • Hold counter-participant utterance.

  • Concatenate participant and counter-participant utterance and classify satisfaction.

  • The next utterance from participant will be given the satisfaction from the concatenated utterance from the previous utterances. reflecting the satisfaction at that given point in time.

Parameters:

  • annotated_dialog_file – Annotated dialogue json file.

  • participant_to_learn – Which participant we want to create a template on.

  • satisfaction_classifier – SatisfactionClassifier.

Returns:

Dictionary with intents as keys and lists of corresponding templates.