Emotion Recognition in Conversation Based on a Dynamic Complementary Graph Convolutional Network

Zhenyu Yang,Xiaoyang Li,Yuhu Cheng,Tong Zhang,Xuesong Wang

DOI: https://doi.org/10.1109/taffc.2024.3360979

IF: 13.99

2024-01-01

IEEE Transactions on Affective Computing

Abstract:Emotion recognition in conversation (ERC) is a widely used technology in both affective dialogue bots and dialogue recommendation scenarios, where motivating a system to correctly recognize human emotions is crucial. Uncovering as much contextual information as possible with a limited amount of dialogue information is essential for eventually identifying the correct emotion of each sentence. The integration of contextual information using the existing approaches often results in inadequate access to information or information redundancy. Deeply integrating the different knowledge behind utterances is also difficult. Therefore, to address these problems, we propose a dynamic complementary graph convolutional network (DCGCN) for conversational emotion recognition. Our approach uses commonsense knowledge to complement the contextual information contained in utterances and enrich the extracted conversation information. We creatively propose the concept of utterance density to prevent redundancy and the loss of utterance information in context-dependent contextual information modeling cases. An utterance dependency structure is dynamically determined by the utterance density, and the contextual information is fully integrated into each sentence representation. We evaluate our proposed model in extensive experiments conducted on four public benchmark datasets that are commonly used for ERC. The results demonstrate the effectiveness of the DCGCN, which achieves competitive results in terms of well-known evaluation metrics. Our code is available at https://github.com/Tars-is-a-robot/Conversational-emotion-recognition.git.

computer science, cybernetics, artificial intelligence

What problem does this paper attempt to address?

This paper attempts to address several key issues in Emotion Recognition in Conversation (ERC):

1. **Insufficient Context Information Acquisition**: Existing methods often lead to insufficient or redundant information when integrating context information. This makes it difficult for the system to extract enough context information from limited conversation data to correctly recognize the emotion of each utterance.

2. **Difficulty in Integrating Common Sense Knowledge**: Although introducing common sense knowledge can enhance emotion recognition, effectively combining this knowledge with conversation content remains a challenge.

3. **Dynamic Context Modeling**: Existing methods usually use a fixed window size when constructing conversation graph structures, which limits the ability to dynamically adjust context information, resulting in insufficient information acquisition.

To address these issues, the paper proposes a Dynamic Complementary Graph Convolutional Network (DCGCN). The main contributions of this method include:

1. **Proposing the Utterance Density Graph (UDG)**: By calculating the speaking density of each speaker in the conversation, the edges in the graph structure are dynamically adjusted to achieve effective propagation of context information.

2. **Integrating Common Sense Knowledge**: Utilizing common sense knowledge generated from external knowledge bases (such as ATOMIC), deep integration with conversation content is achieved through deep graph convolution, enriching the conversation information.

3. **Experimental Validation**: Extensive experiments were conducted on four commonly used ERC datasets, and the results show that DCGCN achieves significantly better performance than other models on multiple evaluation metrics.

In summary, this paper aims to improve the accuracy and robustness of emotion recognition in conversation by enhancing methods for context information acquisition and common sense knowledge integration.