UNLP-2025

Social Media Manipulation Detection in Ukrainian Telegram Posts

UNLP-2025 Shared Task: Detecting Social Media Manipulation

Purdue University Fort Wayne, 2025

Multi-label classification and span-level detection of manipulation techniques in Ukrainian Telegram posts. This work focused on identifying propaganda and manipulation techniques using transformer-based models with explainability and failure-mode analysis.

Task Description

  • Multi-label Classification: Identify multiple manipulation techniques present in social media posts
  • Span Detection: Locate specific text spans where manipulation occurs
  • Context: Ukrainian Telegram posts in high-stakes disinformation environment

Approach

  • RoBERTa-based models with rationale-guided training
  • Emphasis on explainability and interpretable predictions
  • Systematic failure-mode analysis to understand model limitations
  • Cross-lingual transfer learning techniques

Key Findings

  • Achieved competitive performance on manipulation technique classification
  • Identified systematic challenges in span-level detection for code-mixed text
  • Demonstrated importance of explainability in high-stakes NLP applications

Skills & Tools

Python, Transformers, RoBERTa, XLM-R, Multi-label Classification, Span Detection, Explainable AI