@inproceedings{mandravickaite-etal-2025-exploring,
    title = "Exploring Hate Speech Detection Models for {L}ithuanian Language",
    author = "Mandravickait{\.{e}}, Justina  and
      Rimkien{\.{e}}, Egl{\.{e}}  and
      Petkevi{\v{c}}ius, Mindaugas  and
      Songailait{\.{e}}, Milita  and
      Zaranka, Eimantas  and
      Krilavi{\v{c}}ius, Tomas",
    editor = "Calabrese, Agostina  and
      de Kock, Christine  and
      Nozza, Debora  and
      Plaza-del-Arco, Flor Miriam  and
      Talat, Zeerak  and
      Vargas, Francielle",
    booktitle = "Proceedings of the The 9th Workshop on Online Abuse and Harms (WOAH)",
    month = aug,
    year = "2025",
    address = "Vienna, Austria",
    publisher = "Association for Computational Linguistics",
    url = "https://preview.aclanthology.org/landing_page/2025.woah-1.18/",
    pages = "206--218",
    ISBN = "979-8-89176-105-6",
    abstract = "Online hate speech poses a significant challenge, as it can incite violence and contribute to social polarization. This study evaluates traditional machine learning, deep learning and large language models (LLMs) for Lithuanian hate speech detection, addressing class imbalance issue via data augmentation and resampling techniques. Our dataset included 27,358 user-generated comments, annotated into Neutral language (56{\%}), Offensive language (29{\%}) and Hate speech (15{\%}). We trained BiLSTM, LSTM, CNN, SVM, and Random Forest models and fine-tuned Multilingual BERT, LitLat BERT, Electra, RWKV, ChatGPT, LT-Llama-2, and Gemma-2 models. Additionally, we pre-trained Electra for Lithuanian. Models were evaluated using accuracy and weighted F1-score. On the imbalanced dataset, LitLat BERT (0.76 weighted F1-score) and Multilingual BERT (0.73 weighted F1-score) performed best. Over-sampling further boosted weighted F1-scores, with Multilingual BERT (0.85) and LitLat BERT (0.84) outperforming other models. Over-sampling combined with augmentation provided the best overall results. Under-sampling led to performance declines and was less effective. Finally, fine-tuning LLMs improved their accuracy which highlighted the importance of fine-tuning for more specialized NLP tasks."
}