CredenceAI at SemEval-2026 Task 10: A Span-Consistency Network with Cross-Marker Attention for Conspiracy Marker Extraction

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

Summary

CredenceAI introduced a Span-Consistency Network (SCN) for extracting five character-level conspiracy marker types (Actor, Action, Effect, Evidence, Victim) from English social media text, evaluated under overlap-based Macro F1. This system addresses challenges like fragmented spans, ignored inter-marker dependencies, and severe class imbalance. Its architecture includes a Span Consistency Layer (SCL) for coherent boundary formation, Cross-Marker Attention (CMA) to model co-occurrence patterns, and Span Count Regularization (SCR) to align token probabilities with expected discrete spans. Built on DeBERTa-v3-large and trained with a recall-biased Tversky loss, the ensembled system achieved a Macro F1 of 0.24 on the official SemEval-2026 Task 10 test set, securing second place. Ablation studies highlighted SCR's importance for low-frequency and long-span markers.

Key takeaway

For NLP engineers tackling complex span extraction tasks with class imbalance or overlapping entities, CredenceAI's Span-Consistency Network offers a robust architectural blueprint. You should consider integrating components like a Span Consistency Layer for boundary coherence, Cross-Marker Attention for inter-entity dependencies, and especially Span Count Regularization to prevent prediction collapse and improve performance on rare or long spans. This can significantly enhance the quality of your extracted entities.

Key insights

A Span-Consistency Network effectively extracts conspiracy markers by integrating span coherence, inter-marker dependencies, and span count regularization.

Principles

• Propagate span-level confidence for coherent boundaries.
• Model co-occurrence patterns between entity types.
• Align soft token probabilities with expected discrete spans.

Method

The approach combines a Span Consistency Layer, Cross-Marker Attention, and Span Count Regularization on a DeBERTa-v3-large base, trained with recall-biased Tversky loss and ensembled across five stratified folds.

In practice

• Use Span Count Regularization for low-frequency spans.
• Apply Cross-Marker Attention to model entity co-occurrence.
• Employ recall-biased Tversky loss for class imbalance.

Topics

• Conspiracy Marker Extraction
• Span-Consistency Network
• Cross-Marker Attention
• Span Count Regularization
• DeBERTa-v3-large
• SemEval-2026
• Natural Language Processing

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

Related on AIssential

• Explicit Evidence Grounding via Structured Inline Citation Generation — LLMs excel at document retrieval but struggle with precise evidence span identification for faithful inline citations.
• AttackPathGNN: Cross-function vulnerability detection in smart contracts using state interference graphs and conjunction pooling — Cross-function smart contract vulnerabilities require reasoning over inter-function relationships, not just single-function patterns.
• 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.
• Harnessing Structural Context for Entity Alignment Foundation Models — Explicitly harnessing structural context in both encoding and decoding improves transferable Entity Alignment.
• Evidence Graph Consistency in Retrieval-Augmented Generation: A Model-Dependent Analysis of Hallucination Detection — Evidence Graph Consistency (EGC) uses structural relationships to detect RAG hallucinations, but its diagnostic direction varies significantly across LLM families.
• 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.

Open in AIssential →

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