Natural Language Generation (NLG)

The NLG components in Dialoguekit are currently template-based. The basic template-based generation works by extracting templates from a set of training data. The templates are the user and agent utterances with the annotations removed. This version can be found here: dialoguekit.nlg.nlg_template.TemplateNLG.

An extended version of the basic template-based implementation which can perform conditional language generation is dialoguekit.nlg.nlg_conditional.ConditionalNLG. Using a conditional may be useful in cases were other attributes then the intent and annotations may be of interest. Such as the users satisfaction. This can then be used to selecting templates that have a different tone based on the conditional, in this case the satisfaction score.

These two are used in similar ways. The only difference is that the ConditionalNLG can use a conditional in the dialoguekit.core.annotated_utterance.AnnotatedUtterance.metadata as long as the conditional value is a number.

To start using these NLG classes a template needs to be generated. Two methods for doing this are provided:

• dialoguekit.nlg.template_from_training_data.build_template_from_instances()

• dialoguekit.nlg.template_from_training_data.extract_utterance_template()

Usage Example

from dialoguekit.core.annotation import Annotation
from dialoguekit.core.intent import Intent
from dialoguekit.nlg.nlg import NLG
from dialoguekit.nlg.template_from_training_data import (
    extract_utterance_template,
)

template = extract_utterance_template(
    annotated_dialogue_file=ANNOTATED_DIALOGUE_FILE_PATH,
)
nlg = NLG(response_templates=template)

response = nlg.generate_utterance_text(intent=Intent("COMPLETE"))

ConditionalNLG

The ConditionalNLG class contains two additional parameters to the generate_utterance_text() method.

• conditional: str

• conditional_value: Number

These fields are used to find the utterance that has the closest conditional_value to the utterance metadata field with the key conditional. Note that this necessitates the need of a larger template with multiple values for the conditional.

Custom NLG

Any custom NLG component might be implemented by inheriting from dialoguekit.nlg.nlg_abstract.AbstractNLG.