Comparative Analysis of Encoder-Based and Decoder-Based Architectures for Automatic Conspiracy Theory Identification

This study evaluates the performance of encoder-based and decoder-based architectures for the Automatic Con-spiracy Theory Identification (ACTI) task, focusing on Subtask A, which involves detecting conspiratorial content in Telegram posts. I compared RoBERTa-based encoder models and OpenAI GPT-2 decoder models; both fine-tuned with uniform hyperparameters. My encoder-only model achieved an Fl-score of 0.8211, while the decoder-only model reached 0.7464, both significantly outperforming the Kaggle baseline (0.5107) but falling short of the EVALITA winners' score of 0.8572. The encoder model showed stronger performance in identifying conspiratorial content with fewer false positives, whereas the decoder model struggled with an increased number of false negatives. The results underscore the encoder-based model's effectiveness in this context and contribute to understanding the strengths and limitations of these architectures for conspiracy theory detection