Translated vs Non-Translated Method for Multilingual Hate Speech Identification in Twitter

Abstract

Nowadays social media is often misused to spread hate speech. Spreading hate speech is an act that needs to be handled in a special way because it can undermine or discriminate other people and cause conflict that leading to both material and immaterial losses. There are several challenges in building a hate speech identification system; one of them is identifying hate speech in multilingual scope. In this paper, we adapt and compare two methods in multilingual text classification which are translated (with and without language identification) and non-translated method for multilingual hate speech identification (including Hindi, English, and Indonesian language) using machine learning approach. We use some classification algorithms (classifiers) namely Support Vector Machine (SVM), Naive Bayes (NB), and Random Forest Decision Tree (RFDT) with word n-grams and char n-grams (character n-grams) as feature extraction. Our experiment result shows that the non-translated method gives the best result. However, the use of non-translated method needs to be reconsidered because this method needs more cost for data collection and annotation. Meanwhile, translated without language identification method give a poor result. To address this problem, we combine translated method with monolingual hate speech identification, and the experiment result shows that this approach can increase the multilingual hate speech identification performance compared to translate without language identification. This paper discusses the advantages and disadvantages for all method and the future works to enhance the performance in multilingual hate speech identification.

Towards multidomain and multilingual abusive language detection: a survey

“…Based on Table 4, most studies implemented transformerbased architecture to deal with abusive language detection in a cross-lingual setting. However, we also observe some works that exploited a traditional machine learning approach, such as logistic regression [6,10,135], linear support vector machines [92,94], and support vector machines [59]. They used multilingual language representation or simple translation tools (to translate the data training to the target languages) for the knowledge sharing between languages.…”

“…[120] Traditional models Experimented with the use of machine translation tools to translate the training data to the target language and exploited a wide range of traditional models including SVM, naïve Bayes, and random forest. [59] Neural based Proposed a joint-learning architecture based on LSTM coupled with features from HurtLex to transfer knowledge between domains and languages. [94] Transformer based Proposed multichannel architecture based on BERT model, which learns the task sequentially in three languages: source languages, English, and Chinese.…”

“…A study by Basile et al [10] proposed to use TF-IDF representation of bleached characters n-grams. Other studies simply translated the training data to the target language and used the word n-grams feature representation [6,59,92,94]. Meanwhile, most neural-based models were coupled by multilingual word embedding models, including Facebook MUSE (Multilingual FastText) [6,89,92,94] and Babylon Embeddings [89].…”

Cross-Lingual Few-Shot Hate Speech and Offensive Language Detection Using Meta Learning

“…Ousidhoum et al [46] presented the first multilingual multi-aspect hate speech analysis dataset in English, French, and Arabic tweets and evaluated several multilingual multi-task learning approaches for the identification of hate in a multilingual setting. Ibrohim and Budi [54] investigated the effect of the machine translation approach in multilingual hate speech detection in Hindi, English, and Indonesian, by comparing classifiers trained with/without translating samples. Ranasinghe and Zampieri [44] employed a cross-lingual contextual word embeddings model, XLM-R, to transfer knowledge from a rich-resourced language, English, to a lower-resource language (i.e., Bengali, Hindi, or Spanish) to predict offensive content in less-resourced languages.…”

Naive Bayes modification for intrusion detection system classification with zero probability

“…Naïve Bayes algorithm is a classification algorithm that is quite good and is often used in various studies [11]- [13]. This algorithm can be used for simple classification with fixed Y variable and also for text classification [14]- [16]. Laga and Sarno [17] showed that Naïve Bayes gave the best accuracy from other classification methods, such as KNN, SVM, and random forest.…”