Citation

Dewani A, Memon MA, Bhatti S. J. Big Data 2021; 8(1): 160.

Copyright

(Copyright © 2021, Holtzbrinck Springer Nature Publishing Group)

DOI

10.1186/s40537-021-00550-7

PMID

34956818

PMCID

PMC8693595

Abstract

Social media have become a very viable medium for communication, collaboration, exchange of information, knowledge, and ideas. However, due to anonymity preservation, the incidents of hate speech and cyberbullying have been diversified across the globe. This intimidating problem has recently sought the attention of researchers and scholars worldwide and studies have been undertaken to formulate solution strategies for automatic detection of cyberaggression and hate speech, varying from machine learning models with vast features to more complex deep neural network models and different SN platforms. However, the existing research is directed towards mature languages and highlights a huge gap in newly embraced resource poor languages. One such language that has been recently adopted worldwide and more specifically by south Asian countries for communication on social media is Roman Urdu i-e Urdu language written using Roman scripting. To address this research gap, we have performed extensive preprocessing on Roman Urdu microtext. This typically involves formation of Roman Urdu slang- phrase dictionary and mapping slangs after tokenization. We have also eliminated cyberbullying domain specific stop words for dimensionality reduction of corpus. The unstructured data were further processed to handle encoded text formats and metadata/non-linguistic features. Furthermore, we performed extensive experiments by implementing RNN-LSTM, RNN-BiLSTM and CNN models varying epochs executions, model layers and tuning hyperparameters to analyze and uncover cyberbullying textual patterns in Roman Urdu. The efficiency and performance of models were evaluated using different metrics to present the comparative analysis.

RESULTS highlight that RNN-LSTM and RNN-BiLSTM performed best and achieved validation accuracy of 85.5 and 85% whereas F1 score was 0.7 and 0.67 respectively over aggression class.

Language: en

Keywords

Big data; Cyberbullying; Advanced preprocessing; Deep learning; Hate speech detection