Morphological Feature Extraction for Fine-Grained Sorani Kurdish Dialect Identification: A Hybrid Transformer-Linguistic Approach
EACL AbjadNLP (Oral) / 2026
Other authors: +2
As reported, approximately 6 million people in Iraq and Iran speak in Sorani Kurdish, which exhibits substantial regional variation but lacks computational resources for dialect identification. We present the first fine-grained sub-dialect classification system for six Sorani varieties namely, Sulaymaniyah, Erbil, Iranian Sorani, Ardalani, Babani, and Mukriani. This investigation combines cross-lingual contextual embeddings (XLM-RoBERTa) with morphological features derived from explicit linguistic rules, including 24 patterns capturing verb prefixes, pronominal clitics, and definite markers. The suggested morphology-augmented XLM-R model has been trained on a unified dataset of 16,409 sentences without manual annotation, and achieves 91.91% accuracy, outperforming pure transformers (91.79%) and traditional machine learning baselines (SVM 86.41%). Key ablation studies reveal that morphological features serve as effective regularizers for geographically proximate dialects.
AjamiMorph: Zero-Annotation Morphological Discovery for Hausa Ajami via Multi-Method Consensus
EACL AbjadNLP (Oral) / 2026
Other authors: +2
Hausa Ajami (Hausa written in Arabic script) remains severely under-resourced for computational morphology. We present AjamiMorph, a zero-annotation framework that discovers morphemes through consensus among three unsupervised methods, namely, Byte Pair Encoding (BPE), transition-based boundary detection using Pointwise Mutual Information (PMI), and computational linguistics based Distributional Affix Mining (DAM). Using a Hausa Ajami Bible corpus consisting of 637,414 tokens, AjamiMorph identifies 1,611 high-confidence morphemes, achieving 99.9% coverage. The inventory exhibits a linguistically realistic distribution (66.0% stems, 22.6% suffixes, 11.4% prefixes) and recovers 77.8% of known Hausa affixes. A permutation test that shuffles method assignments (preserving per-method selection sizes) confirms that the observed agreement is above-chance; chi-square remains as a secondary check. A lightweight 5-gram LM comparison (characters vs. consensus morphemes) provides an extrinsic signal. We also report negative results for script-driven Arabic assumptions and LLM-first annotation. This work provides the first unsupervised morpheme inventory for Hausa Ajami and demonstrates consensus as a robust strategy for zero-resource morphology.
HCAT-Net a Novel Hierarchical Cross-Attention Transformer Network with Enriched Balanced Ordinal Loss for EEG Emotion Classification
IEEE CIACON / 2025
Other authors: +3
Recognizing human emotions from electroencephalography (EEG) signals is crucial for applications like affective computing and healthcare, yet faces challenges due to the complex, non-stationary nature of EEG data and the inherent ordinal structure of many emotional states. This study introduces the Hierarchical Cross-Attention Transformer Network (HCAT-Net), a novel deep learning architecture specifically designed for improved ordinal emotion classification from 1D EEG time-series. HCAT-Net integrates 1D Residual Convolutional Neural Network blocks with a two-stage hierarchical Transformer encoder, enhanced by Rotary Positional Embeddings (RoPE), to effectively capture dependencies across multiple time scales. The study presents a cross-attention mechanism for improved context integration across Transformer stages and introduces a balanced ordinal loss combining cross-entropy with scaled MAE to preserve label order. Evaluated rigorously on a benchmark EEG emotion dataset, HCAT-Net achieves 99.8% test accuracy, demonstrating its effectiveness. Detailed architectural descriptions and formulations are provided, presenting HCAT-Net as a powerful new approach.
The Glass Policy: A Quantum-Inspired RL Framework for Brittle Mastery in Upskilling
QuantumKol / 2026
Machine learning has mostly always depended on large, static datasets, thereby creating a barrier in dynamic, data-scarce environments. This work addresses the cold-start problem by proposing a quantum-inspired Reinforcement Learning framework, QABACUS (Quantum Adaptive Budget-Aware Cold-start Upskilling System). The agent policy is modeled as a variational quantum-inspired state and frames personalized employee training as a constrained RL problem, wherein an agent learns an optimal policy to maximize skills while adhering to strict rule-based budgetary constraints. We use an Actor-Critic agent to optimize the parameters of this policy wavefunction, exploring a superposition of potential upskilling strategies. The experiments introduce objective metrics, including Return on Investment and policy entropy, to evaluate performance. The optimal QABACUS configuration demonstrates clear improvement in a deterministic setting with an 85% success rate and fastest convergence to a stable, low-entropy policy. However, the study reveals a critical limitation: this optimized policy undergoes catastrophic failure analogous to policy decoherence under environmental stochasticity, highlighting the sim-to-real gap caused by deterministic overfitting. These results validate the quantum-inspired RL approach as a strong paradigm for data-scarce decision support, while identifying generalization under uncertainty as the key challenge for practical deployment.
