<![CDATA[The widespread dissemination of hate speech and offensive content on social media platforms has become a critical societal issue, highlighting the need for reliable automated detection systems. This study proposes a hybrid approach that leverages frozen embeddings from the pre-trained language model cardiffnlp/twitter-roberta-base-offensive as a high-level semantic feature extractor, combined with the CatBoost gradient boosting algorithm as the final classifier. The proposed method was evaluated on the HASOC 2021 English dataset through six experimental scenarios and compared with a TF-IDF baseline using CatBoost's default hyperparameters. Experimental results demonstrate that the proposed approach achieved a Macro F1-score of 0.7924 for the binary classification task (Task 1A) and 0.6113 for the multiclass classification task (Task 1B), outperforming the TF-IDF baseline, which achieved scores of 0.7724 and 0.5798, respectively. The proposed system demonstrated a clear performance improvement and achieved results comparable to those of the top-ranked teams on the official HASOC 2021 leaderboard, while avoiding the computational cost associated with fine-tuning large pre-trained language models.]]>
Copyrights © 2026
