|
Mohaouchane, Mourhir & Nikolov (2019)
|
Remove non-Arabic letters, special characters, emoticons, diacritics, punctuation, elongated words. Letter ormalization and tokenization |
15,050 YouTube comments Classes: neutral, offensive, non-offensive |
CNN-LSTM Rec = 83.46%, Acc = 87.27% Prec = 83.89%, F1 = 83.65% CNN F1 = 84.5%, Acc = 87.84%, Prec = 86.10% |
The study is limited to YouTube comments and may not apply to other platforms. Future research should include Arabic text in Latin alphabet and dialects, and identify other objectionable content on social platforms. |
|
Abu Farha & Magdy (2020)
|
Remove unknown characters, diacritics, punctuation Remove elongation, URLs Normalization |
OSACT4: 10,000 tweets Classes: Hate-Speech, Non hate-speech, Offensive, Non-offensive |
Hate-speech ( MTL: F1 = 76%) Offensive language (MTL-S-N: F1 = 87.7%) |
Risk of mistake transmission with external sentiment data. Need for lexicon augmentation. Multitask learning settings for improved outcomes. Future research will incorporate sentiment data for hate speech and objectionable language. |
|
Faris et al. (2020)
|
Remove non-Arabic characters, symbols, numbers, punctuation, hashtags, web addresses, stop words, diacritics. Normalization. Tokenization |
Private/3,696 tweet Classes: Hate, normal |
CNN+ LSTM Used AraVec N-grams + SG model with 50 epochs Acc = 66.564% Rec = 79.768% Pre = 68.965% F1 = 71.68% |
The model’s robustness decreases with smaller datasets, and detecting Arabic hate speech requires larger benchmark datasets and a larger Arabic lexicon, necessitating further research into deep learning methods. |
|
Duwairi, Hayajneh & Quwaider (2021)
|
Removing numbers, English characters, links, hashtags, use mentions, emoji’s, and punctuations. Normalizing stripping Arabic diacritics. |
ArHS dataset/9,833 tweets Classes: Binary (normal, hate) Ternary (normal, hate, abusive) Multi-class (normal, misogyny, religious, abusive, racism, and discrimination) |
Using the ArHS dataset: In the binary CNN F1 = Pre = Acc = 81 In the ternary CNN and BiLSTM-CNN Acc = 74 In the multi-class CNN-LSTM, BiLSTM-CNN Acc = 73 |
Previous research on hate speech detection has largely overlooked its complexity, suggesting future studies should expand the ArHS dataset’s application to NLP tasks, explore deep learning models, and consider demographic information. |
|
Al-Hassan & Al-Dossari (2021)
|
Remove punctuations, repeated characters, @username, URLs, hashtags.) Normalization. |
11,000 tweets Classes: Religious, Racism, Sexism, General hate speech, Not hate speech |
CNN + LTSM Pre = 72% Rec = 75% F1 = 73%. |
The study of Arabic tweets faces challenges due to limited data, linguistic diversity, and subjective categorization. Future plans include expanding the dataset, developing real-time hate speech classification methods, exploring alternative text representation techniques, and using advanced hardware. |
|
Mazari & Kheddar (2023)
|
Keeping only Arabic and French characters, deleting numbers, diacritics, elongation, repeated letters more than twice, unknown Unicode, and extra spaces. Substitution of URLs, user mentions, and emoticons with tags and hashtags with separated words. |
14, 150 comments from Facebook, YouTube, and Twitter classes: Hate speech (HS), Cyberbullying (CB), Offensive language (OF). |
Bi-GRU Acc = 73.6% F1 = 75.8% |
The study of toxic Arabic and hate speech lacks available tools and datasets, and modern NLP methods like BERT, GPT-2, and GPT-3 were not used. Future research should focus on expanding dialectal datasets to include other Arab languages and determining effective word embeddings and methods for dialect toxicity detection. |
|
Al-Ibrahim, Ali & Najadat (2023)
|
Removing Punctuations, links, Numbers, Non-Arabic Characters, Repeated Letters, Arabic Diacritics, spaces and blank lines. Normalization |
Private/15,000 tweets Classes: hate/non-hate |
Improved Bi-LSTM Acc = 92.20% F1 = 92% |
The dataset is accurate but limited to Twitter. Future expansion should include religious, ethnic, political, and hate speech categories, and other platforms such as YouTube. |
|
Hassan et al. (2020)
|
Remove diacritics, words that contain non-Arabic characters, punctuation. Repeated characters replaced with only one. |
OSACT4/10,000 tweets Classes: Hate/non hate Offensive/not offensive |
Combination of SVMs, CNN-BiLSTM, and M-BERT. Subtask A: F1 = 90.51% Subtask B: F1 = 80.63% |
The distribution of hate speech data is uneven, with concerns about vocabulary and word ambiguity. Skewed data makes diagnosing system breakdowns difficult. Future studies aim to identify errors and improve systems by augmenting hate speech data. |
|
Haddad et al. (2020)
|
Removing diacritics, punctuation, non-Arabic characters, emoticons, and stop words, replacing some tokens by their Arabic words. Reducing repeated letters and elongated words. Normalization. |
OffensEval 2020/10,000 tweets, and a YouTube comments dataset Classes: offensive, Inoffensive, Hate Speech, Not Hate Speech |
Bi-GRU-ATT Offensive (F1 = 85.9% Acc = 91%) Hate speech (F1 = 75% Acc = 95%) |
The authors suggest using LSTM models instead of GRU models for improved hate speech detection, as it can better distinguish between hate speech and offensive language, and suggest future research on this approach. |
|
Guellil et al. (2021)
|
Removing diacritics, and Transliteration may be applied to convert Arabizi to Arabic script. Tokenization Normalization. |
5,000 YouTube Comments. Classes: hateful, non-hateful |
CNN F1 = 86% |
The authors aim to increase the 5,000-comment corpus to 10,000 automatically, using automated methods to handle Arabizi, detect Arabic, French, and English comments, and improve categorization performance through transliteration and language identification. They plan to link hate speech identification to sentiment analysis and include a transliteration technique. |
|
Al Anezi (2022)
|
Remove irrelevant data, diacritics, symbols, special characters, and emoji’s. |
4,203 comments/not mentioned platforms Classes: Binary (negative/positive) Multi-class (religion, race, gender, violent/offensive, bullying, normal positive, and normal negative) |
DRNN-1 Binary Acc = 99.73% DRNN-2 Multi-class Acc = 95.38% |
The dataset for hate speech in Arabic could be expanded in size and class count, and machine learning techniques could be improved for higher accuracy. Future plans include creating a prototype system for real-time hate speech monitoring and processing, which can be integrated into social media platforms. |
|
Alshalan & Al-Khalifa (2020)
|
Remove hashtag, stop words and replacing emoji with textual description, punctuation, whitespaces, diacritics, non-Arabic Characters Lemmatization. Normalization. |
Private dataset (GHSD)/9316 Tweetsc Classes: hateful, abusive and normal |
CNN Pre = 81% F1 = 79% Rec = 78% AUROC = 89% |
Hate speech detection is challenging due to tweets, Arabic dialects, and MSA linguistic variations. Future plans include expanding the dataset, adding multi-labels, and testing models on other abusive language datasets, potentially improving classifier performance. |
|
Orabe et al. (2020)
|
Removing repeated consecutive characters, some stopwords, some punctuation marks, and keeping emoji’s within the texts |
10,000 tweets classes: offensive (OFF) and not-offensive (NOT OFF) |
BI-GRU-ATT Pre = 88% F1 = 85% Rec = 83% Acc = 91% |
The dataset has an imbalance between offensive and non-offensive samples, with 81% being non-offensive. To rectify this, the authors plan to use oversampling/undersampling approaches to gather more offensive samples and enhance the dataset with additional abusive and inoffensive tweets. |
|
Ahmed, Maher & Khudhur (2023)
|
Removing special characters, URLs, words in foreign languages, emoji’s, and extra spaces. Tokenization. |
32,000 comments/Aljazeera.net Classes: clean, obscene, and rude. |
Three-layer LSTM Acc = 92.75%. |
High performance but limited to specific domains. Future work focuses on improving deep learning algorithms for Arabic text cyberbullying detection, sentiment analysis, and synthetic methods to protect minors from cybercrime on all social media platforms. |
|
Mousa et al. (2024)
|
Cleaning, normalization, Farasa segmentation, and tokenization. |
13,000 tweets multiclasses: racism, bullying, insult, obscene language, and non-offensive content. |
1D-CNN cascaded with RBF with 88.2% F1-score |
This research’s limitations include high code complexity from cascaded models, lengthy training periods due to large dataset size, and the need for multiple machine learning model combinations. Future work focuses on improving the cascaded model’s performance using models with lower computational complexity. |
