EUVIMEDEUVIMED entdecken
Open Access · CC BY 4.0

Implicit-Emotion Recognition Model Based on Content–Style Decoupling and Conditional Fusion

Yi Zhang, Junqing Zhu, Hua Zhao

Electronics · 08.07.2026 · 15 (14) · 3002

DOI: 10.3390/electronics15143002 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Electronics, online erschienen am 08.07.2026. Verfasst von Yi Zhang und 2 weiteren Autor:innen. Die Arbeit nennt 40 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: Shandong Provincial Natural Science Foundation, National College Student Innovation and Entrepreneurship Training Program.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Implicit-emotional expressions are common in college students’ social media posts, where literal meanings may contradict underlying emotions. Bidirectional Encoder Representations from Transformers (BERT)-based models face two coupled challenges in this setting: feature confusion between semantic content and expression style, and coarse-grained feature fusion. We propose CSD-IFRN (Content–Style Disentanglement with Conditional Fusion for Implicit Emotion Recognition Network), which disentangles content and style using dual non-shared BERT encoders, combines gradient-reversal adversarial training with a content–style orthogonality regularizer (Lorth), and applies conditional layer normalization (CLN) for adaptive fusion. On a dedicated dataset of 11,154 triple-annotated texts, averaged over five random seeds, CSD-IFRN achieves 88.58% accuracy and 88.43% macro-F1, improving over BERT-base-chinese by 6.99 points and over the strongest SOTA baseline by 2.61 points. The main gains remain significant after Holm correction (p < 0.01) and also hold on a public benchmark. A frozen style probe trained on content features falls to 50.40% balanced accuracy, close to chance level, supporting effective disentanglement. Among fusion strategies, CLN achieves the best accuracy with low seed-to-seed variance. These results suggest that CSD-IFRN can provide an auxiliary signal for university mental-health monitoring, rather than a clinical diagnostic tool.

Förderung

  • Shandong Provincial Natural Science Foundation
  • National College Student Innovation and Entrepreneurship Training Program

Zitieren

APA

Zhang, Y., Zhu, J., & Zhao, H. (2026). Implicit-Emotion Recognition Model Based on Content–Style Decoupling and Conditional Fusion. Electronics, 15(14), 3002. https://doi.org/10.3390/electronics15143002

Vancouver

Zhang Y, Zhu J, Zhao H. Implicit-Emotion Recognition Model Based on Content–Style Decoupling and Conditional Fusion. Electronics. 2026;15(14):3002. doi:10.3390/electronics15143002

BibTeX
@article{Zhang2026,
  author = {Zhang, Yi and Zhu, Junqing and Zhao, Hua},
  title = {Implicit-Emotion Recognition Model Based on Content–Style Decoupling and Conditional Fusion},
  journal = {Electronics},
  year = {2026},
  volume = {15},
  number = {14},
  pages = {3002},
  doi = {10.3390/electronics15143002}
}
Kostenlos starten →