2021/05/25 by Djamila Romaissa Beddiar, Beddiar, Djamila Romaissa, Md Saroar Jahan +3 · 1 citation
Computer Science · #Computation and Language (cs.CL) #FOS: Computer and information sciences #Hate Speech and Cyberbullying Detection
paper · pdf · doi:10.48550/arxiv.2106.04681
openalex publication_date 2021/05/25 · openalex created_date 2022/07/25 · openalex updated_date 2026/07/28
With proliferation of user generated contents in social media platforms,\nestablishing mechanisms to automatically identify toxic and abusive content\nbecomes a prime concern for regulators, researchers, and society. Keeping the\nbalance between freedom of speech and respecting each other dignity is a major\nconcern of social media platform regulators. Although, automatic detection of\noffensive content using deep learning approaches seems to provide encouraging\nresults, training deep learning-based models requires large amounts of\nhigh-quality labeled data, which is often missing. In this regard, we present\nin this paper a new deep learning-based method that fuses a Back Translation\nmethod, and a Paraphrasing technique for data augmentation. Our pipeline\ninvestigates different word-embedding-based architectures for classification of\nhate speech. The back translation technique relies on an encoder-decoder\narchitecture pre-trained on a large corpus and mostly used for machine\ntranslation. In addition, paraphrasing exploits the transformer model and the\nmixture of experts to generate diverse paraphrases. Finally, LSTM, and CNN are\ncompared to seek enhanced classification results. We evaluate our proposal on\nfive publicly available datasets; namely, AskFm corpus, Formspring dataset,\nWarner and Waseem dataset, Olid, and Wikipedia toxic comments dataset. The\nperformance of the proposal together with comparison to some related\nstate-of-art results demonstrate the effectiveness and soundness of our\nproposal.\n