UCSC NLP at SemEval-2026 Task 10: Boundary-Aware Span Extraction and RoBERTa Classification for 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

UCSC NLP developed systems for SemEval-2026 Task 10 (PsyCoMark), focusing on conspiracy marker extraction (Subtask 1) and document-level conspiracy detection (Subtask 2). For Subtask 1, the team employed multi-label span classification over candidate spans, utilizing IoU≥0.95 positive labeling, hard-negative sampling, and containment-based non-maximum suppression with boundary-aware span representations. Subtask 2 involved an independent sequence classifier, enhanced with label smoothing and a stratified train–validation split. Analysis revealed robust detection for entity-like roles such as Actor and Victim, while abstract roles like Action, Effect, and Evidence proved sensitive to boundary criteria. On the official test set, their systems achieved 7th place in Subtask 1 with a 0.2251 macro F1 score and 12th place in Subtask 2 with a 0.7694 weighted F1 score.

Key takeaway

For NLP engineers developing conspiracy detection systems or participating in similar span extraction tasks, you should prioritize robust boundary-aware span representations, especially when dealing with abstract concepts. Consider that entity-like roles are more reliably identified than abstract ones, suggesting a need for specialized handling or data augmentation for Action, Effect, and Evidence roles. Your system's performance on such nuanced tasks can significantly benefit from techniques like hard-negative sampling and label smoothing.

Key insights

UCSC NLP's systems combine boundary-aware span extraction and RoBERTa classification for conspiracy detection in SemEval-2026 Task 10.

Principles

• Entity-like roles (Actor, Victim) are robustly detectable.
• Abstract roles (Action, Effect, Evidence) are sensitive to boundary criteria.

Method

Marker extraction uses multi-label span classification with IoU≥0.95 labeling, hard-negative sampling, and containment-based NMS; document classification uses a sequence classifier with label smoothing.

In practice

• Employ IoU≥0.95 for positive span labeling.
• Use hard-negative sampling for training.
• Apply containment-based NMS for span refinement.

Topics

• SemEval-2026
• Conspiracy Detection
• Span Extraction
• RoBERTa Classification
• Multi-label Classification
• Natural Language Processing

Best for: Research Scientist, AI Scientist, NLP 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.
• Introducing Span Categorization in Prodigy and spaCy — Span Categorization (spancat) in spaCy extracts and classifies long, overlapping text spans, providing explicit control and confidence scores.
• Explicit Evidence Grounding via Structured Inline Citation Generation — LLMs excel at document retrieval but struggle with precise evidence span identification for faithful inline citations.
• Extending an Ensemble Baseline with Corpus-Based Graph Features for Portuguese Pun Detection — Graph-based features can augment pun detection, but their utility depends on specific lexical relation encoding.
• Structured Sentiment Analysis in Brazilian Portuguese: An Exploratory Study Using BERTimbau — Span-level sentiment extraction provides a viable baseline for Structured Sentiment Analysis in Brazilian Portuguese.
• FineREX: Fine-Tuned NER-RE for Human Smuggling Knowledge Graphs — Domain-specific fine-tuning of LLMs significantly enhances knowledge graph construction quality and efficiency for complex, jargon-heavy legal texts.

Open in AIssential →

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