Impact of Encoder Architecture and Input Features on Dialogue Act Classification: A Comparative Study of RNN Encoders

Salah-Eddine EL MOUSLIH, Imad Boudroua

Abstract: This paper examines the impact of encoder architecture and input features on dialogue act classification, an important task in dialogue systems. We conduct several experiments comparing the performance of different recurrent neural network encoders. These include a GRU encoder initialized with BERT weights that considers only the previous utterance as context, and the same model with speaker-level embeddings. We also compare two variants of a bi-directional LSTM encoder for dialogue act classification: one that takes multi-utterance conversations of BERT pooled outputs with and without speaker-level embeddings, and another that averages the LSTM layer outputs. Our findings indicate that incorporating a bi-directional LSTM encoder with BERT's pooled representation improves classification performance significantly.