The Importance of Context in Intent Classification: A Comparative Study of Encoder-Decoder Architectures

Clara Le Gallic-Ach, Coni Soret

Abstract: This study presents a benchmark of various encoder and decoder architectures for intent classification in dialogue systems, using the DailyDialog corpus. The role of context in classification accuracy is explored, with a particular focus on the importance of capturing dynamic structures of context in real-world applications. Our results demonstrate that including context significantly improves classification performance, and that the choice of decoder architecture is important both for their architecture and the level of context they use. We also demonstrated that analyzing accuracy according to the context - past, none, and full - provides valuable insights into the impact of context on real-world applications.