HAM-GNN: A Hierarchical Attention-Based Multi-Dimensional Edge Graph Neural Network for Dialogue Act Classification

Dialogue act (DA) analysis is crucial for developing natural conversational systems and dialogue generation. Modelling DA labels at the utterance-level requires contextual and speaker-aware understanding, especially for conversational agents handling Japanese dialogues. In this study, we propose a Hierarchical Attention- based Multi-dimensional Edge Graph Neural Network (HAM-GNN) to effectively model DA labels by capturing speaker interconnections and contextual semantics. Specifically, long short-term memory networks (LSTMs) first encode contextual information within a dialogue window. We then construct a context graph by aggregating neighbouring utterances and apply a graph attention network (GAT) to model speaker interactions with multi-dimensional edges. To prevent incorrect edge definitions from completely deactivating connections during training, we initialize soft edges for nominally unconnected nodes with a small non-zero value. Moreover, to avoid loss of contextual information from localized graph construction, we utilize a convolutional Transformer (Conformer) to build residual connections. Subsequently, a gated graph convolutional network (GatedGCN) selects salient utterances for DA classification. Finally, multi-level representations are merged and fed to dense layers for classification. We evaluate our HAM-GNN model on the Japanese DA dataset (JPS-DA) and the English Switchboard DA dataset (SWDA). Results show our method outperforms baselines on JPS-DA and achieves competitive performance on SWDA. The graph-based architecture effectively encodes utterance semantics and speaker relationships for DA prediction in conversational systems.