Dialog Act Classification with BERT Models

Beatriz Farah, Serine Mechide

Abstract: The identification of Dialog Acts (DA) through sequence labeling systems is an important part of Spoken Dialog (SD) understanding. Nowadays DA recognition has gained attention given its importance in Chatbot training, since understanding the role that a user's message plays in a message is crucial in order to respond in an appropriate and helpful way. In this work, we perform DA classification adapted to SD, which we implement using 2 BERT models (Bidirectional Encoder Representations from Transformers). We evaluate the models on 3 DA databases from the SILICONE benchmark, and we compare the models' performances by obtaining their accuracy, their average training time and their loss.