KUET at StanceNakba Shared Task: StanceMoE: Mixture-of-Experts Architecture for Stance Detection

2026-04-01

AI Summary

This study addresses the challenge of modeling diverse linguistic signals in implicit target stance detection, where unified representations often fail to capture heterogeneous expressive patterns. To this end, the authors propose StanceMoE, a context-enhanced Mixture-of-Experts architecture built upon fine-tuned BERT. StanceMoE introduces, for the first time in stance detection, a Mixture-of-Experts mechanism comprising six specialized expert modules designed to capture semantic orientation, salient lexical cues, clause-level focus, phrasal patterns, framing signals, and contrastive discourse shifts. A context-aware gating network dynamically weights and fuses these expert outputs based on input characteristics. Evaluated on the StanceNakba 2026 Subtask A dataset, StanceMoE achieves a macro-F1 score of 94.26%, significantly outperforming existing baselines and BERT variants, thereby demonstrating its strong adaptive capacity in modeling heterogeneous stance expressions.

Abstract

Actor-level stance detection aims to determine an author expressed position toward specific geopolitical actors mentioned or implicated in a text. Although transformer-based models have achieved relatively good performance in stance classification, they typically rely on unified representations that may not sufficiently capture heterogeneous linguistic signals, such as contrastive discourse structures, framing cues, and salient lexical indicators. This motivates the need for adaptive architectures that explicitly model diverse stance-expressive patterns. In this paper, we propose StanceMoE, a context-enhanced Mixture-of-Experts (MoE) architecture built upon a fine-tuned BERT encoder for actor-level stance detection. Our model integrates six expert modules designed to capture complementary linguistic signals, including global semantic orientation, salient lexical cues, clause-level focus, phrase-level patterns, framing indicators, and contrast-driven discourse shifts. A context-aware gating mechanism dynamically weights expert contributions, enabling adaptive routing based on input characteristics. Experiments are conducted on the StanceNakba 2026 Subtask A dataset, comprising 1,401 annotated English texts where the target actor is implicit in the text. StanceMoE achieves a macro-F1 score of 94.26%, outperforming traditional baselines, and alternative BERT-based variants.

Problem

• stance detection
• actor-level stance
• geopolitical actors
• linguistic signals
• implicit targets

Innovation

• Mixture-of-Experts
• stance detection
• context-aware gating
• linguistic signals
• BERT-based architecture