Intents Classification for Neural Text Generation

Sirine Louati, Jérémie Stym-Popper

Abstract:

Sequence labelling tasks like Dialog Acts (DA) and Emotion/Sentiment (E/S) are very important in spoken dialog systems. In fact, they allow distinguishing different dialog acts and emotions from different conversations.
In this work, we propose an intent classifier that allows to identify different labels such as communicative intent or dialog acts from a given conversation. We evaluate this classifier on the SILICONE benchmark introduced by Chapuis et al. [2020]. Our experiments show that using an encoder with the BERT model enables to build a strong intent classifier achieving good performances. All our numer- ical experiments and codes are located here : https://github.com/Jeremstym/NLP_intent_class