Citation

Uddin, Abu Zahid Md Jalal and Hossain, Md. Mithun and Akib, A. S. M. Ahsanul Sarkar and Mridha, M. F. and Hossen, Md. Jakir (2026) Routed Expert-Based Emotion Detector for Multi-Label Emotion Classification. IEEE Open Journal of the Computer Society, 7. pp. 876-888. ISSN 2644-1268

Text IEEE Xplore Full-Text PDF_5.pdf - Published Version Restricted to Repository staff only Download (1MB)

Abstract

Multi-label emotion detection aims to identify multiple emotions that may co-occur in short, informal texts, such as tweets. Real-world data present challenges due to overlapping emotional cues, complex inter-label dependencies, and severe imbalance between frequent and rare emotions. Existing transformer-based approaches typically rely on fixed thresholds and a single inductive bias, such as conditional prediction, graph reasoning, or flat classification, which often lead to overfitting on majority labels and poorly calibrated probability estimates. To address these limitations, we propose ROUTED (Routed Expertbased Emotion Detector), a routed mixture-of-experts framework that adaptively selects the most appropriate reasoning strategy for each input instance. ROUTED integrates three complementary experts: a label-graph expert that models emotion co-occurrence using a PPMI-based graph, a classifier-chain expert that captures directional dependencies among labels, and a lightweight MLP expert for order-free scoring. A compact router assigns instance-specific mixture weights through a temperature-scaled softmax with entropy-based load balancing to prevent expert collapse. Training combines binary cross-entropy and focal loss to mitigate class imbalance, incorporates Laplacian graph regularization to enforce correlation consistency, and applies per-label threshold tuning for calibrated inference. Experiments on SemEval-2018 Task 1 (EC, English) and REN-CE-cps (Chinese) demonstrate consistent macro-F1 improvements over strong baselines, particularly for rare emotions, while providing interpretable attention patterns and expert-usage analyzes that reveal the model’s adaptive reasoning behavior.

Item Type:	Article
Uncontrolled Keywords:	Multi-label, emotion detection, mixture-of-experts,
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA190-194 Management of engineering works
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Rosnani Abd Wahab
Date Deposited:	05 Jun 2026 06:50
Last Modified:	05 Jun 2026 06:50
URII:	http://shdl.mmu.edu.my/id/eprint/16046

Downloads

Edit (login required)