CuriosAI at SemEval-2026 Task 10:Hybrid approaches to conspiracy span extraction and conspiracy detection

2026-06-25 · Source: Paper Index on ACL Anthology · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, quick

Summary

CuriosAI's system for SemEval-2026 Task 10 addresses Conspiracy Marker Extraction and Conspiracy Detection. For marker extraction, the system employs multi-label token classification using a bidirectional transformer, specifically DeBERTa-v3-large, to predict overlapping spans. Feature-based and LLM-based alternatives did not outperform this encoder baseline. For conspiracy detection, CuriosAI compared various heterogeneous models, including transformer fine-tuning, lexical classifiers, embedding-based models, and LLM-based refinement. The study found that models optimal during development did not consistently generalize best. Logit-level ensembling ultimately achieved the strongest test performance, with an F1 score of 0.7620. These results underscore the value of bidirectional token modeling for effective span extraction and calibration-aware ensembling for robust detection.

Key takeaway

For NLP Engineers developing systems for misinformation detection, you should prioritize bidirectional transformer architectures like DeBERTa-v3-large for accurate span extraction. When building your detection models, do not solely rely on development-optimal performance; instead, integrate calibration-aware, logit-level ensembling to achieve more robust and generalizable results. This approach can significantly improve your system's F1 score, as demonstrated by the 0.7620 performance in SemEval-2026 Task 10.

Key insights

Bidirectional transformers and calibration-aware ensembling are key for robust conspiracy span extraction and detection.

Principles

• Bidirectional token modeling enhances span extraction.
• Development-optimal models may not generalize best.
• Calibration-aware ensembling improves detection robustness.

Method

For marker extraction, use multi-label token classification with a bidirectional transformer. For detection, compare heterogeneous models and apply logit-level ensembling.

In practice

• Implement DeBERTa-v3-large for token classification.
• Test logit-level ensembling for model fusion.
• Evaluate generalization beyond development metrics.

Topics

• Conspiracy Detection
• Span Extraction
• Transformers
• DeBERTa-v3-large
• Ensemble Learning
• SemEval-2026

Best for: Research Scientist, AI Scientist, NLP Engineer, Machine Learning Engineer

Related on AIssential

• mdok-style at SemEval-2026 Task 10: Finetuning LLMs for Conspiracy Detection — Data augmentation and self-training can effectively finetune LLMs for specialized text classification tasks with limited data.
• KITE: A Tri-Modal Transformer Integrating Text, Images, and Knowledge Graphs for Fake News Detection — KITE integrates text, images, and knowledge graphs via a tri-modal transformer for enhanced fake news detection and interpretability.
• Explicit Evidence Grounding via Structured Inline Citation Generation — LLMs excel at document retrieval but struggle with precise evidence span identification for faithful inline citations.
• A Multimodal Framework for Financial Fake News Detection for Brazilian Portuguese — Multimodal analysis, integrating text and image-extracted text, significantly improves financial fake news detection.
• Robust Dual-Signal Fusion: Hybrid Neuro-Symbolic Gating with Compressed Chain-of-Thought Refinement for Irony Detection in Social Media Texts — Hybrid neuro-symbolic fusion with compressed Chain-of-Thought effectively overcomes LLM literalism for zero-shot irony detection.
• The Geometry of Truth: Detecting LLM Hallucinations with Geometric Algebra — Type II hallucinations are context contradictions, dangerous due to their plausibility and standard similarity metric blindness.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Paper Index on ACL Anthology.