EUVIMEDEUVIMED entdecken
Crossref

A Cross-Domain Knowledge-Enhanced Framework for Fine-Grained Emotion-Cause Pair Extraction with Optimized Pre-training

B. Kranthi Kumar, Prof. S. Aquter Babu

Research Digest on Engineering Management and Social Innovations · 24.11.2025 · 41

DOI: 10.46647/xvszab68 ↗

Kurz gefasst

Fachartikel in der Zeitschrift Research Digest on Engineering Management and Social Innovations, online erschienen am 24.11.2025. Verfasst von B. Kranthi Kumar und 1 weiteren Autor:innen. Die Arbeit wurde bisher 1-mal zitiert (Crossref).

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Textual Emotion Cause Extraction (TECE) has emerged as a critical task in natural language processing, enabling systems to identify not only emotions expressed in text but also their underlying causes. Traditional Emotion Cause Extraction (ECE) methods suffer from two fundamental limitations: dependency on pre-annotated emotions and separate handling of emotion recognition and cause extraction, neglecting their interdependent nature. While Emotion-Cause Pair Extraction (ECPE) addresses these issues through joint detection, current state-of-the-art approaches face persistent challenges including error propagation in two-step pipelines, coarse-grained clause-level extraction, limited cross-domain generalization, insufficient knowledge integration, and suboptimal pre-training strategies. This paper proposes a novel cross-domain knowledge-enhanced framework that advances ECPE through four key innovations: (1) fine-grained span-level extraction that identifies precise emotion and cause phrases rather than entire clauses, enabling aspect-level analysis; (2) cross-domain adaptation mechanisms combining adversarial training and domain-invariant feature learning for effective transfer across news, social media, reviews, and conversational domains; (3) multi-source knowledge injection architecture integrating linguistic patterns, discourse relations, commonsense reasoning from knowledge bases (ConceptNet, ATOMIC), and emotion lexicons through knowledge-aware graph attention networks; and (4) optimized pre-training strategy combining domain-adaptive masked language modeling with emotion-aware contrastive learning and discourse relation prediction. The proposed framework employs a unified end-to-end architecture with three core modules: a Knowledge-Optimized Pre-training Layer using customized objectives for TECE, a Span-Level ECE Module with multi-task BIO tagging for emotion/cause/aspect spans, and a Knowledge-Enhanced Reasoning Layer using dependency-based attention and knowledge-guided GAT over clause graphs enriched with commonsense concepts. Cross-domain adaptation is achieved through domain discriminators and parameter-efficient domain-specific adapters. Comprehensive experiments on benchmark datasets (ECPE-news, enhanced English ECPE, cross-domain corpora) demonstrate that our framework achieves significant improvements: F1 scores of 74.6% on standard ECPE tasks (4.2% improvement over state-of-the-art), 71.8% on cross-domain scenarios (6.8% improvement), and 78.3% on span-level extraction (5.1% improvement). Ablation studies validate that knowledge enhancement contributes 2.8% F1 gain, cross-domain adaptation 3.2%, and optimized pre-training 2.1%. The fine-grained extraction capability enables precise aspect-level emotion-cause analysis, while knowledge enhancement significantly improves handling of implicit causality (8.4% improvement on implicit examples). This work advances the state-of-the-art in emotion-cause analysis and establishes practical foundations for deployment in mental health monitoring, customer experience management, social media analytics, and intelligent dialogue systems. The proposed framework addresses critical real-world requirements for robustness across domains, fine-grained actionable insights, and effective handling of complex implicit emotion-cause relationships.

Zitieren

APA

Kumar, B. K., & Babu, P. S. A. (2025). A Cross-Domain Knowledge-Enhanced Framework for Fine-Grained Emotion-Cause Pair Extraction with Optimized Pre-training. Research Digest on Engineering Management and Social Innovations, 41. https://doi.org/10.46647/xvszab68

Vancouver

Kumar BK, Babu PSA. A Cross-Domain Knowledge-Enhanced Framework for Fine-Grained Emotion-Cause Pair Extraction with Optimized Pre-training. Research Digest on Engineering Management and Social Innovations. 2025:41. doi:10.46647/xvszab68

BibTeX
@article{Kumar2025,
  author = {Kumar, B. Kranthi and Babu, Prof. S. Aquter},
  title = {A Cross-Domain Knowledge-Enhanced Framework for Fine-Grained Emotion-Cause Pair Extraction with Optimized Pre-training},
  journal = {Research Digest on Engineering Management and Social Innovations},
  year = {2025},
  pages = {41},
  doi = {10.46647/xvszab68}
}
Kostenlos starten →