GUNLP at SemEval-2026 Task 10: Psycholinguistic Conspiracy Marker Extraction and Detection (PsyCoMark)
Year: 2026
Type: conference-paper
Abstract: This paper presents the Georgetown University NLP (GUNLP) system for SemEval 2026 Task 10: Psycholinguistic Conspiracy Marker Extraction and Detection.Specifically, our system addresses Subtask 2 by classifying conspiratorial beliefs in Reddit posts.Our approach uses COVID-Twitter-BERT v2 (CT-BERT-v2) (Müller et al., 2023), a BERT-large model pretrained on COVID-19 social media data.We integrate this model into a multi-task learning framework with a dual-head architecture on the [CLS] token to jointly optimize conspiracy classification and emotion prediction.To enrich the training set, we use GPT-5 to generate chain-of-thought emotion annotations and apply paraphrase-based data augmentation.This effectively doubles the training corpus to approximately 8,600 samples.We evaluate two input configurations: text only, and text with emotion annotations.The emotion-aware configuration achieves the strongest performance, reaching an F1 score of 0.54 (Yes: 0.655, No: 0.752, Can't Tell: 0.222) on the official development set, with an official test set score of 0.34 1 * Equal contribution.
Authors: Rojin Ziaei, Mahsa Khoshnoodi, Nazli Goharian
Institution: Georgetown University
Language: English
Subfield: Sociology and Political Science
Field: Social Sciences
Domain: Social Sciences
Open Access status: hybrid
OpenAlex ID: W7166838657