In recent various approaches have been proposed by researchers towards hate speech detection for major OSNs like Twitter, Facebook, Reddit, Wikipedia, YouTube etc. We described some of the approaches as follows.
Ross, B., et al. have measured the reliability of the hate speech and to what extent in accordance with subjective ratings [18]. Davidson, T. et al. have used a crowd-sourced hate speech lexicon to collect the hate speech from the tweets. They have the multiclass technique to distinguish between different categories of the hate speeches [19]. Badjatiya, P. have used deep learning framework for hate speech detection from tweets. They defined this problem by classifying tweets into categories like racist, sexist or neither. They have used the benchmark dataset of annotated tweets of 16K and stated that the deep learning techniques outperformed the char/n-gram techniques [20].
Gao, L. et al. have used logistic regression and neural network models for the hate speech detection. They have provided the corpus of the hate speech dataset. They have stated that both models have performed well on the benchmark dataset and achieved better results in comparison with the baseline classifiers [21]. Founta, A. M. et al. have proposed incremental and iterative methods for the detection of abusive language on the social media platforms like Facebook and Twitter. According to them, the proposed methodology working better for the reduction but robust labels to characterize the abusive tweets [22]. Zhang, Z. et al. have targeted to identify the characteristics of the tweets like race, and religion. They have stated that the hate speech detection is a challenging task due to unique, discriminative features. They have proposed the Deep Neural Network for the features extraction and to capture the semantics of the hate speech [23].
Gröndahl, T. et al. have suggested that the data and labeling is more important than the accurate hate speech detection models [24]. Mishra, P. et al. have addressed the problem of obfuscation of words by users to evade detection model. They have designed the model for embeddings for unseen words. They have stated that their approach achieved significant improvement in the detection of hate speech on Twitter and Wikipedia datasets [25].
Kshirsagar, R. et al. have presented neural network based approach for classification of hate and non-hate Twitter speech basically in racist and sexist. They have used three datasets namely, Sexist/Racist (SR), HATE and HAR. They have used word embedding and pooling features to train the deep neural network [26].
Qian, J., et al. presented two large fully-labeled datasets collected from Gab and Reddit. They evaluated the datasets in order to better understand common intervention tactics and to investigate the effectiveness of common automatic response generation methods [27]. Mansourifar, H. et al. have collected significant dataset from the clubhouse. They have analyzed the dataset stastically using the Google Perspective Scores. According to them, the Perspective Scores outperforms the Bag of Words andWord2Vec textual features [28].
Gautam, A. et al. have presented the dataset related to MeToo movement. They have manually annotated the dataset for five different linguistic aspects like, relevance, stance, hate speech, sarcasm, and dialogue acts [29]. Silva, L. have analyzed the targets of hate speech in online social media. The have collected the traces of the two social media like Twitter and Whisper. According to them, their approach identifies the hate speeches and provides the directions for prevention and detection approaches [30].
Salminen, J. et al. have manually labeled the posts from YouTube and Facebook videos. They have created taxonomy of different types of targets and trained machine learning classifiers to automatically detect the online hate speeches. They have conducted experiments using machine learning classifiers, like Logistic Regression, Decision Tree, Random Forest, Adaboost, and Linear SVM, to generate a multi-class, multi-label classification model that automatically detects and classifies hateful comments. They have achieved best performance using Linear SVM with an average F1 score of 0.79 using TF-IDF features [31]. ElSherief, M. et al. have presented the comparative study of hate speech instigators and target users on Twitter. According to them, hate instigators target more popular and high profile Twitter users [32].