Developing Relationships: A Heterogeneous Graph Network with Learnable Edge Representation for Emotion Identification in Conversations

Emotion recognition in conversations (ERC) aims to predict the emotion of utterances. Modeling context dependencies is the critical challenge of the task. Existing efforts in ERC are mainly based on the sequence and graph models. The graph models can better capture structured information than the sequence models. Unfortunately, there are few suitable aggregation strategies for ERC models based on high-dimensional edge features. Moreover, the adjustment of edge representation in graph-based models has been ignored for a long time. Based on this, we propose a learnable edge message-passing model based on a heterogeneous dialog graph. The model first calculates the attention weights between utterance nodes and between nodes and edges separately and then learns contextual utterance representations through these learnable edge representations. Additionally, we conducted our experiment on four public datasets and achieved advanced results.