Transfer Learning via Lexical Relatedness: A Sarcasm and Hate Speech Case Study (arXiv)

Social networks continue to have trouble identifying indirect hate speech, such as irony, sarcasm, and insinuation. The current study examines whether pre-training with sarcasm improves the detection of hate speech, both explicit and implicit. Two training procedures were evaluated on CNN+LSTM and BERT+BiLSTM models using datasets from ETHOS, Reddit sarcasm, and the Implicit hatred Corpus: one that used sequential transfer learning across sarcasm, implicit hatred, and explicit hate, and another that trained only on sarcasm. The results demonstrated that pre-training with sarcasm considerably improved BERT+BiLSTM performance, including precision on implicit samples and recall, AUC, and F1-score on ETHOS. The results imply that sarcasm-aware algorithms are more adept at detecting subtle hate speech.

https://arxiv.org/abs/2508.16555