NUST PsyAI at SemEval-2026 Task 10: Parameter-Efficient RoBERTa for Conspiracy Detection and Character-Level Marker Extraction

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

Summary

The NUST PsyAI system participated in SemEval-2026 Task 10 (PsyCoMark), focusing on document-level conspiracy detection and character-level psycholinguistic marker extraction from Reddit discourse. The system employed a parameter-efficient RoBERTa-large model finetuned with LoRA. For conspiracy detection, the RB-DET-LoRA approach achieved a weighted F1 score of 0.79 on the development set and 0.76 on the test set, ranking 8th overall. In the character-level marker extraction task, the system ranked 7th, achieving an Overlap F1 of 0.16 on the development set and 0.21 on the test set. The extraction methodology explored both unified multi-type BIO and decomposed per-type schemes, with the latter proving more effective by mitigating cross-label interference and improving boundary consistency. The analysis highlighted that detection benefits from contextual semantic modeling, while extraction faces limitations due to sparse supervision and sensitive boundary evaluation.

Key takeaway

For NLP Engineers developing systems for psycholinguistic marker extraction, you should consider the inherent challenges of sparse supervision and boundary sensitivity. While parameter-efficient RoBERTa with LoRA is effective for document-level detection, your extraction models may benefit from decomposed per-type schemes to mitigate cross-label interference. Prioritize robust data annotation and augmentation strategies for character-level tasks to improve boundary consistency and overall performance.

Key insights

Parameter-efficient RoBERTa with LoRA effectively detects conspiracies, but marker extraction struggles with sparse data and boundary sensitivity.

Principles

• Contextual semantic modeling aids detection.
• Decomposed per-type extraction reduces interference.
• Sparse supervision limits extraction performance.

Method

The NUST PsyAI system finetunes RoBERTa-large using LoRA for parameter-efficient learning. Extraction contrasts unified multi-type BIO with a decomposed per-type setup to improve boundary consistency.

In practice

• Apply LoRA for efficient RoBERTa finetuning.
• Use decomposed schemes for complex entity extraction.
• Prioritize data augmentation for sparse extraction tasks.

Topics

• SemEval-2026
• Conspiracy Detection
• Psycholinguistic Markers
• RoBERTa
• LoRA
• Parameter-Efficient Finetuning
• Reddit Discourse

Best for: Research Scientist, AI Scientist, NLP Engineer

Related on AIssential

• An End-to-End Hybrid Framework for Rumour Detection in Low-Resources Algerian Dialect — A hybrid framework combining transformer embeddings with classical classifiers enables effective rumour detection in low-resource Algerian dialect.
• A 1B humanizer that matches human writing on an AI detector — A 1B-parameter humanizer can generate AI text that evades current AI detectors, matching human writing.
• mdok-style at SemEval-2026 Task 9: Finetuning LLMs for Multilingual Polarization Detection — Finetuning mid-sized LLMs with QLoRA and augmented data improves multilingual online polarization detection.
• 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.
• 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.
• One Language, Two Scripts: Probing Script-Invariance in LLM Concept Representations — SAE features capture abstract semantic meaning independent of surface-level orthography and tokenization.

Open in AIssential →

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