RoBERTa vs BERT for intent classification

Gabriel Guaquiere, Anna Nguyen

Abstract:

Intent classification is an essential task in nat-
ural language processing, which aims to iden-
tify the intention or purpose behind a user’s ut-
terance. This task has become increasingly im-
portant in the development of conversational
agents and chatbots, as they need to under-
stand user requests to provide relevant and
accurate responses. In this paper1, we will
look at which algorithm, between BERT and
RoBERTA, seems more adapted to the pre-
diction of dialogue act (DA) or sentiment and
emotion (S/E). We will use the SILICONE
dataset which seems to fit the task. It contains
corpora including an utterance and the asso-
ciated DA or S/E. We observe that RoBERTa
outperforms BERT in prediction, especially
for DA. For the mrda dataset, it even man-
ages to reach an accuracy of 89%. For the pre-
diction of S/E, its performance is better than
BERT, nevertheless its predictivity rate is low.