Residual-Based Graph Convolutional Network for Emotion Recognition in Conversation for Smart Internet of Things

Area of Science:

• Artificial Intelligence
• Natural Language Processing
• Machine Learning

Background:

• Emotion recognition in conversation (ERC) is vital for user-centric Internet of Things devices.
• Current deep learning methods often use recurrent networks and attention, with recent models incorporating graph networks for utterance relationships.
• A gap exists in extracting detailed intra-utterance features before sequence processing.

Purpose of the Study:

• To propose a novel Residual-based Graph Convolution Network (RGCN) for improved emotion recognition in conversation.
• To develop a new loss function that enhances model effectiveness by considering edge weights.
• To fully exploit both intra-utterance and inter-utterance features for more accurate ERC.

Main Methods:

• A Residual Network (ResNet)-based module for detailed intra-utterance feature extraction.
• A Graph Convolutional Network (GCN)-based module for inter-utterance feature extraction, capturing conversational context.
• Integration of ResNet and GCN within the proposed RGCN framework.
• Introduction of a novel loss function incorporating edge weights.

Main Results:

• The proposed RGCN method significantly outperforms existing state-of-the-art methods in emotion recognition tasks.
• The ResNet-based extractor effectively captures context within individual utterances.
• The GCN-based extractor successfully leverages inter-utterance relationships for condensed feature representation.

Conclusions:

• The RGCN model offers a superior approach to emotion recognition in conversation by effectively combining intra- and inter-utterance feature extraction.
• The proposed loss function contributes to the improved performance by weighting edge importance.
• This work advances the field of conversational AI by providing a more accurate and nuanced method for understanding user emotions.