Kshitij Rajput, Raghav Kapoor, Puneet Mathur, Hitkul, Ponnurangam Kumaraguru, and Rajiv Ratn Shah, Springer (2020).


With the phenomenal increase in the penetration of social media in linguistically diverse demographic regions, conversations have become more casual and multilingual. The rise of informal code-switched multilingual languages makes it tough for automated systems to monitor instances of hate speech, which are further intelligently disguised through the use of spelling variations, code-mixing, homophones, homonyms, and the absence of sophisticated grammar rules. Machine transliteration can be employed for converting the code-switched text into a singular script but poses the challenge of the semantical breakdown of the text. To overcome this drawback, this chapter investigates the application of transfer learning. The CNN-based neural models are trained on a large dataset of hateful tweets in a chosen primary language, followed by retraining on the small transliterated dataset in the same language. Since transfer learning can act as an effective strategy to reuse already learned features in learning a specialized task through cross-domain knowledge transfer, hate speech classification on a large English corpus can act as source tasks to help in obtaining pre-trained deep learning classifiers for the target task of classifying tweets translated in English from other code-switched languages. Effects of the different types of popular word embeddings and multiple supervised inputs such as the LIWC, the presence of profanities, and sentiment are carefully studied to derive the most representative combination of input settings that can help achieve state-of-the-art hate speech detection from code-switched multilingual short texts on Twitter.