JIA at SemEval-2026 Task 10: A Dual-Track System with BERT-based Encoders and LLMs for Conspiracy Analysis
Year: 2026
Type: conference-paper
Abstract: The proliferation of conspiracy theories on social media platforms like Reddit poses a severe challenge to public discourse, and the accurate identification of these causal narratives relies on the fine-grained capture of their intricate narrative structures.This study proposes a comprehensive monitoring framework that integrates BERT-based discriminative models with Large Language Models (LLMs) to jointly optimize conspiracy detection and psycholinguistic marker extraction.We systematically evaluated the performance of several architectures, including DistilBERT, BERT-Base, DeBERTa-V3, RoBERTa, and the instruction-tuned Qwen2.5 series.Experimental results demonstrate that Qwen2.5-14B(Fullshot) achieves a peak Weighted F1-score of 0.80 in the detection task.Furthermore, marker extraction remains a significant challenge; although the fine-tuned LLM leads in identifying "Actors" (F1 of 0.144), its performance in semantically ambiguous categories such as "Evidence" and "Effect" highlights a pronounced generalization gap.
Author: Jiayue Zhu
Institution: University of Tübingen
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166887747