The intent classifier transforms a natural language utterance into a named intent and a set of named parameters. It uses an ML classifier to infer the intent name and an entity recognizer to extract named entities as parameters. These components are both represented as ML pipelines.

The intent classifier also maintains a tokenizer pipeline for converting utterances into a list of tokens. This pipeline is executed first, and its results are run through the classifier/recognizer pipelines.

Classification results are returned as a tuple of , where intents is the map of names to classification probabilities, and parameters is a name->value map. Intent names, parameter names and parameter values are all strings. Probabilities are all floats that should sum to 1.

Example: pipeline = %{

tokenizer: [{:ptb_tokenizer, []}],
classifier: [{:count_vectorizer, []},
             {:linear_classifier, [c: 2.0, probability?: true]}],
recognizer: [{:crf_tagger, []}],

} x = [

"you have four pears",
"three hundred apples would be a lot"

] y = [

{"intent_1", ["o", "o", "b_count", "b_fruit"]},
{"intent_2", ["b_count", "i_count", "b_fruit", "o", "o", "o", "o"]}

] classifier = Penelope.NLP.IntentClassifier.fit(%{}, x, y, pipeline)

{intents, params} = Penelope.NLP.IntentClassifier.predict_intent(

classifier,
%{},
"I have three bananas"

)