Scholarlink

<p>Emotion recognition in conversations(ERC) aims to identify the emotional state of an utterance based on contextual information. Existing problems: (1) Emotions being unstable and implicit, where implicit emotions may not be directly accessible through multimodal information. (2) Using only simple graph networks for modality modeling ignores the variability and complexity of relationships between modalities, although it reflects the dependencies between modalities and the coherence of contextual information. (3) Uneven sample sizes and similar emotion labels lead to poor ERC performance. To address these challenges, this paper proposes an Intentional Tendency-based Dynamic Heterogeneous Graph Network for ERC(IT-DHGN). Firstly, intentional tendencies are introduced to understand the intention behind each utterance, thereby providing additional emotional features for recognition. Secondly, when constructing a heterogeneous graph, intentional tendencies, text, audio and video are treated as heterogeneous nodes. The Dynamic Heterogeneous Graph Network module(DHGN) explores the heterogeneity between different data objects by integrating them, using the graph structure to dynamically capture dependencies and complementary information between modalities. Lastly, in the Contrast-Enhanced Optimization module(CEO), heterogeneity-based and homogeneity-based contrast enhancement strategies are employed to address the issue of indistinguishable samples with similar emotion labels in ERC. We conducted experiments on two conversation sentiment datasets and performed extensive analysis to validate the effectiveness of the IT-DHGN model.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Intentional tendency-based dynamic heterogeneous graph network for emotion recognition in conversations

  • 期刊论文

  • 发表日期：

    2025年2月17日

查看全文
  • Xinyi Gan,
  • Xianying Huang,
  • Shihao Zou

摘要

Emotion recognition in conversations(ERC) aims to identify the emotional state of an utterance based on contextual information. Existing problems: (1) Emotions being unstable and implicit, where implicit emotions may not be directly accessible through multimodal information. (2) Using only simple graph networks for modality modeling ignores the variability and complexity of relationships between modalities, although it reflects the dependencies between modalities and the coherence of contextual information. (3) Uneven sample sizes and similar emotion labels lead to poor ERC performance. To address these challenges, this paper proposes an Intentional Tendency-based Dynamic Heterogeneous Graph Network for ERC(IT-DHGN). Firstly, intentional tendencies are introduced to understand the intention behind each utterance, thereby providing additional emotional features for recognition. Secondly, when constructing a heterogeneous graph, intentional tendencies, text, audio and video are treated as heterogeneous nodes. The Dynamic Heterogeneous Graph Network module(DHGN) explores the heterogeneity between different data objects by integrating them, using the graph structure to dynamically capture dependencies and complementary information between modalities. Lastly, in the Contrast-Enhanced Optimization module(CEO), heterogeneity-based and homogeneity-based contrast enhancement strategies are employed to address the issue of indistinguishable samples with similar emotion labels in ERC. We conducted experiments on two conversation sentiment datasets and performed extensive analysis to validate the effectiveness of the IT-DHGN model.