Description
This PR brings some improvement to the intent classification by leveraging the presence of builtin entities in the training data.
The strategy consists in preprocessing utterances by looking for builtin entities within the utterances, and add markers for each each entity found while removing the builtin entity substring from the utterance. This allows to capture the fact that a builtin entity is present in an utterance, and prevents from learning specific builtin entities occurence such as 42.
Description This PR brings some improvement to the intent classification by leveraging the presence of builtin entities in the training data. The strategy consists in preprocessing utterances by looking for builtin entities within the utterances, and add markers for each each entity found while removing the builtin entity substring from the utterance. This allows to capture the fact that a builtin entity is present in an utterance, and prevents from learning specific builtin entities occurence such as 42.