Abstract
Hate speech poses significant challenges to maintaining healthy online conversations, and automated systems are crucial for its accurate detection and mitigation. In this paper, we (CNLP-NITS-PP) introduce ATLANTIS (Attentive Transformer-LSTM for Named Entity and Token Identification System), a robust model designed to address the pervasive issue of hate speech in online social media platforms. ATLANTIS focuses on hate span identification within sentences labeled as hate speech, framed as a sequence labeling task using BIO notation. Leveraging a Hate dataset enriched with Named Entity Recognition (NER) tags, ATLANTIS effectively identifies hate speech spans within the text by combining contextualized representations and sequential modeling. The empirical results showcase ATLANTIS’s effectiveness in isolating explicit signs of hate from a contextual backdrop, offering a promising solution for creating safer online environments. We achieve a macro F1 score of 0.488 on the public test set and 0.508 on the private test set. This work not only lays the foundation for future advancements in hate-span detection but also emphasizes the importance of model efficiency, interpretability, and expanded training data that encompass diverse linguistic nuances and evolving hate speech trends. Code is available at https://github. com/niyarrbarman/hasoc23
