Improving Dialogue Act Classification by Considering Context: An investigation of dialogues acts using Bert Encoding

Mathieu Garrouty

Abstract: The recognition of Dialog Acts has become a crucial area of research in recent years, particularly with the growth of chat assistants like ChatGPT. The key to generating a conversation that is as natural as possible is to understand the intent behind each message. In this study, we focus on classifying a dataset of multi-turn dialogs between two individuals, with each message labelled according to its Dialog Act. Our objective is to predict the Dialog Act classification. We compare a basic sequence-level model, where the neural network learns from all labelled sequences, with dialog-level models that take into account the context of a dialog. We employ Recurrent Neural Networks, both with and without self-attention mechanisms, and find that our prediction accuracy increases significantly within a comparable training time, highlighting the importance of context for better representation of dialogs in natural language processing.