A Literature Review of Textual Hate Speech Detection Methods and Datasets

Abstract

Online toxic discourses could result in conflicts between groups or harm to online communities. Hate speech is complex and multifaceted harmful or offensive content targeting individuals or groups. Existing literature reviews have generally focused on a particular category of hate speech, and to the best of our knowledge, no review has been dedicated to hate speech datasets. This paper systematically reviews textual hate speech detection systems and highlights their primary datasets, textual features, and machine learning models. The results of this literature review are integrated with content analysis, resulting in several themes for 138 relevant papers. This study shows several approaches that do not provide consistent results in various hate speech categories. The most dominant sets of methods combine more than one deep learning model. Moreover, the analysis of several hate speech datasets shows that many datasets are small in size and are not reliable for various tasks of hate speech detection. Therefore, this study provides the research community with insights and empirical evidence on the intrinsic properties of hate speech and helps communities identify topics for future work.

1. Introduction

Hate speech is a poisonous discourse that can swiftly spread on social media or due to prejudices or disputes between different groups within and across countries [1]. A hate crime refers to crimes committed against a person due to their actual or perceived affiliation with a specific group (https://www.ilga-europe.org/what-we-do/our-advocacy-work/hate-crime-hate-speech, accessed on 1 March 2022) The protected characteristics of Facebook define hate speech as an attack on an individual’s dignity, including their race, origin, or ethnicity. According to Twitter policies, tweets should not be used to threaten or harass others due to their ethnicity, gender, religion, or any other factor. In addition to age, caste, and handicap, YouTube also censors content that promotes violence or hatred toward certain persons or groups. Often, hate speech regarding online radicalization or criminal activities is studied [2]; however, hate speech has also been discussed in other contexts [3].

There is strong motivation to study the automatic detection of hate speech due to the overwhelming online spread of information [4]. The detection of hate speech is crucial to reducing crime and protecting people’s beliefs. This study is especially important in the face of ongoing wars, distorting reality and dehumanizing the attacked Ukrainian nation. Studies have shown an increase in hate speech against China on online social media, especially racist and abusive content accusing people of causing the COVID-19 outbreak. On the other hand, a lower rate of hate speech reduces crime, such as cyberbullying, which significantly affects social tranquility [5], leading to minimal cyber-attacks [6].

Despite the many studies in the field, hate speech is still problematic and challenging. The literature reports that both humans and machine learning models have difficulty detecting hate speech due to the complexity and variety of hate speech categories. Hate is characterized by more extreme behaviors associated with prejudice. Figure 1 shows numerous aspects of hate speech. Hate speech is seen as a layer between aggressive and abusive text; however, all of these share the offensive aspect. On the other hand, sexist, homophobic, and religious hate are relatively different as they target a group of people or a gender. The figure shows that the separation of the concept is complicated, a challenge that has been previously discussed [7]. Specific hate speech definitions are contentious [7]; racist and homophobic tweets, for example, are more likely to be labeled as hate speech than other types of offensive or abusive content. Therefore, there is no way to generalize whether an inflammatory text is hate speech [8]. The expansion of the representation of short documents, such as on Twitter, is a significant issue; they cause additional challenges to the traditional bag-of-words model, resulting in data sparsity due to insufficient contextual information [9]. Furthermore, many datasets [10] may differ due to these classifications and definitions, making it challenging to compare machine learning models. In other words, hate speech notions are all within the umbrella of abusive text, according to Poletto et al. [1].

The literature showed several literature review articles; however, most of them targeted one area of the literature or are relatively old [7,11,12,13,14]. The literature study in [7] was devoted to building a generic metadata architecture for hate speech classification based on predefined score groups using semantic analysis and fuzzy logic analysis. The study in [11] targeted hate speech concerning gender, religion, and race related to cyberterrorism and international legal frameworks; however, the study did not focus on Twitter or datasets for machine learning. Most of the papers cited in [12] are related to the legal literature that defines hate speech for criminal sanctions. The study in [13] was devoted to hate speech geographical aspects, social media platform diversity, and the generic qualitative or quantitative methods used by researchers. To the best of our knowledge, no review has been dedicated to English hate speech dataset analysis. This paper aimed to deeply review hate speech concepts, methods, and datasets to provide researchers with an insight into the latest state-of-the-art studies in hate speech detection.

This study carries a systematic literature review based on the methodology of Tranfield et al. [15]. The method synthesizes and extracts results on evidence-based systematic literature extracted based on four aspects: the research question, review criteria, final literature review, and data extraction and synthesis. The studied topic is heterogeneous in its wide range of methods and homogenous in terms of using textual Twitter hate speech content, but is bounded by a specific research question. Therefore, the methodology of Tranfield et al. is applicable [16].

The central research question is as follows.

RQ: What are the dominant hate speech concepts, datasets, and machine learning methods?

The paper is outlined as follows. Section 2 discusses the research methodology. Results and content analysis are discussed in Section 3. Section 4 discusses research challenges and future directions. Finally, conclusions are illustrated in Section 5.

2. Methodology

This study focuses on machine learning classification models, hate speech datasets, and the most significant features of hate speech models in light of such a widespread phenomenon. We applied a survey and content analysis to retrieve and analyze relevant studies based on the methodology of Tranfield et al. [15]. The methodology of Tranfield et al. [15] is based on four phases: research question formulation, review criteria definition, final literature review criteria, and data extraction and combination. The primary research question in this paper is to identify the dominant approaches to hate speech and its significant datasets that are commonly used by the research community. Following the research question, the review criteria include: (1) selecting a set of hate speech terms combined with machine learning methods. The selection includes terms such as hate, cyberhate, racist language, offensive language, online harm, poisonous speech, religion hate, ethnicity, nationality, race, gender hate, aggressive, abusive, misogyny, sexist speech, homophones, and homophobic language; (2) the following data repositories were consulted: Science Direct, Scopus, Emerald, IEEE Xplore, and Google Scholar; (3) the study was conducted between 2016 and 2021; and (4) papers must be in English with a minimum length of 5 pages. During the data extraction and synthesis phase, the collection of literature was analyzed and then grouped according to the review criteria that were initially put in place.

Figure 2 shows the proposed methodology. The initial analysis process followed the approach of Tranfield et al.; papers that passed the RQ and the initial inclusion criteria were included in this review. The initial analysis is based on the analysis of the paper title, abstract, and introduction. Consequently, papers that did not pass the inclusion criteria were eliminated. Then, the data extraction and synthesis phase included papers from the previous step and comprehensively analyzed them by reading and grouping them into meaningful categories. During this step, the thematic analysis technique was used [17]. The thematic approach is based on the recurring patterns of hate speech detection. With careful reading and analysis, the critical information related to the RQ was coded using the content analysis in each paper. The codes were then converted and grouped into higher-order themes. Finally, after deep understanding and linking knowledge embodied in analyzed papers, a set of crucial challenges was highlighted.

3. Results and Analysis

Following the proposed approach in Figure 2, the following steps were executed. In Step 1, the papers that passed the initial inclusion criteria were analyzed (512 papers). In Step 2, the thematic analysis resulted in the final set of papers (138 papers) analyzed in accordance with the approach of Tranfield et al. Step 3 is discussed in Section 4. The most relevant papers are shown in as shown in Appendix A.

The distribution of relevant papers over the years is illustrated in Figure 3. The figure shows an increase in hate speech detection over the years. Among the leading research conferences that are dedicated to hate speech are the Sem-Eval, HASCOC, and FIRE conference series. The Sem-Eval series conference is one of the most specialist conferences on hate speech. In the Sem-Eval series, Task 6 of SemEval-2019 comprises three sub-tasks [18]: A: detecting offensive/non-offensive speech; B: determining offensive speech (insult/nontarget insult); and C: determining the targeted individual or group demonstrating offensive speech. The Offensive Language Identification Dataset (OLID) (14,100 tweets) [18] is part of SemEval-2019 Task 6. In another conference series, for three languages, namely Hindi, German, and English, the HASOC 2020 track of FIRE2020 [19] contained two goals for detecting hate speech: binary (hate/not hate) and multilabel (hate, offensive, and profanity). The initial challenge of FIRE2020 is to distinguish between non-hate–offensive and hate–offensive information.

The content analysis of this study (Figure 4) shows the distribution of hate speech detection methods. According to the figure and our analysis, the most successful hate speech detection methods use more than one approach for detection (hybrid models). The figure also shows a wide range of shallow TFIDF methods to complex BERT models.

3.1. Machine Learning Hate Speech Models

The thematic structures identified in this study are discussed in these major themes: TFIDF, lexicon-based, traditional, and deep learning methods, and hybrid models.

3.1.1. TFIDF Methods

Figure 4 shows that only 12% of the analyzed papers used traditional machine learning methods based on the TFIDF. The term frequency (TF), paired with Inverse Document Frequency (IDF), is a regularly utilized feature in hate speech identification (TFIDF). The TFIDF considers the statistical aspect of input text and the specificity of terms (or chars or POS tags) in the corpus, giving less frequently used words greater weight. For example, in hate speech tweets, the extensive use of hashtags that signify objectionable material, such as “#BanIslam”, “#whoriental”, and “#whitegenocide”, could indicate hostile speech. Therefore, lexicons, hashtags, and POS tags were used in an association rule system [20] to classify hate speech. Several works have been used in this category, such as the author profiling of the Hate Speech Spreader Detection shared task organized by PAN 2021 [21], char/word-IDF [22,23], and Vader/RoBERTa word embeddings [24], trigram features and POS tags [25,26,27], n-grams [28,29,30,31,32], and emotions [26,33].

3.1.2. Lexicon-Based Methods

Lexicon-based methods depend on using existing keywords (lexicons) fixed by authors or extracted from the literature for hate speech detection. The work of Frenda et al. [34] preprocessed two datasets of misogynistic and sexist tweets over SVM classifiers. Frenda et al. used three datasets: Automatic Misogyny Identification (AMI) IberEval [35], AMI EvalIta [36], and the SRW [28]. A list of 690 lexicon items was added to the TFIDF for characters and words, which was utilized to weight lexical features using Information Gain [34]. The following categories are included in their lexicon: hashtags, vulgarity, abbreviations, femininity, sexuality, and human body. Their model has a precision of 0.76. Their model adopts sentence embeddings, the TFIDF of tweets, Glove Bow over Logistic Regression, and XGboost over the English dataset of the AMI shared problem of EVALITA2018. Their highest performing method was LR, which obtained an accuracy of 0.704. Several other studies have used lexicons in hate speech detection such as the Hate Speech and Offensive Content in Social Media Text (HASOC) detection models [25,37,38]; Jewish and Muslim lexicons [39,40]; hate speech-related migrants and religious minorities of Twitter [41,42,43,44,45,46,47,48,49]; and InferSent and Google [46,50,51].

3.1.3. Deep Learning Methods

The convolutional neural network (CNN) is a class of artificial neural networks which is most commonly applied to analyze visual imagery. However, it has also been used in textual classification. CNNs use filters and a set of pooling layers for hate speech detection tasks [49,52,53,54,55,56,57,58,59,60,61,62,63,64,65,66,67,68,69]. Zimmerman et al. [67] tested Waseem [28] and SemEval 2013 sentiment analysis datasets [68]. Zimmerman et al. employed a CNN structure to represent 50 tokens based on CNN parameters (epochs, weights, and batch size). The best-reported model had three epochs and a batch size of ten for positive and negative classifications, with an F1 average score of 75.98%. Gambäck and Sikdar [69] created a dataset of 6655 tweets and used 4 CNN structures to train it. They used n-grams, Word2Vec, and random vectors created at random. The results (using Word2Vec) showed an F1-score of 78.3%. Pre-trained Glove and FastText models were also employed against women and immigrants at an individual and group level [49]. Similarly, in [53,54], the authors used convolutional neural networks to recognize the profiles of hate speech-based word n-grams. Many alternative CNN architectures were used for various content languages [49,55]. In the study which used the SemEval-2019 dataset, for example, word embedding was utilized to detect hate speech in Spanish and English tweets [55].

A recurrent neural network (RNN) is a neural network that uses internal memory to recall its input. The algorithm family, for example, is utilized in Apple’s Siri and Google’s voice search. Many algorithms, such as LSTMs and GRUs, are included in RNNs. LSTMs were created to solve the problem of vanishing gradients that can occur while training standard RNNs. The GRU is similar to long short-term memory (LSTM) with a forget gate, but lacks an output gate. Hence, it has fewer parameters. Variations of LSTM were also used for many tasks of hate speech detection, such as hate ideologies [70], COVID-19 and the US election [71,72,73], the hate detection of hybrid CNN and RNN [74,75,76], HASOC LSTM models [77,78,79], BiLSTM models [65,80,81], and generative pretrained transformer models [82,83,84,85,86,87,88,89]. Some studies have added additional features to Twitter network analysis such as user profiles for hate speech detection. For example, Founta et al. [90] proposed two-layer RNNs. The unified model was built on tweet characteristics (the Glove technique) and metadata on persons, networks, and content. They used many datasets to test their model, including the cyberbullying dataset [91], the hateful dataset [28], the offensive dataset [25], the sarcasm dataset [92], and the abusive dataset [84]. The model produced variable results depending on the input characteristics and dataset utilized; nevertheless, the RNN and metadata interleaved model were the best, with an average accuracy of 90.2. The dataset of Waseem and Hovy [28] was used by Founta et al. [93] to detect abusive cyberbullying language on Twitter. Their model blends the LSTM architecture with social network analysis to detect hate speech sources. In order to train the LSTM model, they used FastText embeddings. An F1-score of 0.823 was reported. Corazza et al. [94] used the dataset of Waseem and Hovy [28] with the following algorithms: LSTM, GRU, and BiLSTM. N-grams of words; tweet features, such as emojis and emotion lexica; and social network-specific features are among the features employed. They reported an F1-score of 0.823; the best method was LSTM.

Bidirectional Encoder Representations from Transformers (BERTs) is a transformer-based machine learning technique for natural language processing which was pre-trained and developed by Google based on the knowledge extracted from text (vectors) using the surrounding text to establish the context. Several complex models that used BERT are discussed in the literature such as hateful meme challenges [64]; the study of Ron Ahu et al. [95]; ensembles of BERT [96]; Yu’s model of knowledge enhanced vision-language representations [97]; Facebook Hateful Memes (FHM) [98,99]; Liu et al. [100] and BERT over Zampieri [101] three tasks; Caselli et al. [102] who re-trained the BERT for hate speech (HATEBERT), the AbusEval [103], and the HatEval [43] models; Nguyen et al. [104] who employed the BERT model of RoBERTa [105] by training on 80 GB of uncompressed texts of 850 M tweets (16B-word tokens), and BERTweet which outperforms strong baselines RoBERTa-base and XLM-R-base [106]. Caselli et al. [102] used the Reddit Abusive Language English dataset, which contains over one million Reddit messages (43 billion tokens). They compared the following datasets from SemEval 2019: Task 6 OffensEval 2019 [18], the AbusEval [103], and the HatEval [43]. They found that the HATEBERT model was more efficient than the datasets they evaluated, with a 5% increase in precision. Nguyen et al. [104] used the RoBERTa [105] BERT model to train 80 GB of raw texts from 850 M tweets (16 B-word tokens). It was found that BERTweet surpasses the RoBERTa-base and XLM-R-base [106], which are both powerful baselines.

Consequently, RNNs and CNNs have commonly used machine learning methods in deep learning methods; however, research has also shown that hybrid and complex models cover approximately 29% of studied models. For example, Jahan and Oussalah [107] showed that the BERT method covers approximately 33% of studies, followed by LSTM and CNN with 20%.

3.1.4. Hybrid Methods

Hybrid methods combine more than one machine learning method to improve performance, including ensemble models. In addition, hybrid models are considered robust as they incorporate more than one data source, such as the metadata from Twitter. Pitsilis et al. [108] tested an ensemble of RNN classifiers in the datasets of Waseem and Hovy [28]. They used the word frequency vectors and the user sentiment towards each class of hate speech (neutral, racism, or sexism) based on user tweet history. The result showed that the model provided an F1-score of 0.93. Joulin et al. [109] reported the best performance was when using a combination of LSTM + Random Embedding + Boosted Decision Trees (GBDTs) with an F1-score of 0.93 in the dataset of Waseem and Hovy [28] and using Glove embeddings. Paschalides et al. [110] used the dataset of Davidson et al. [25] to develop a new online Twitter hate speech detection system called MANDOLA. The system is based on many ensembles of CNNs and RNNs that automatically learn abstract feature representations depending on many features: TF-IDF vectors and word embeddings. MANDOLA allows users to visualize online results. The approach reported a balanced accuracy of 0.770.

Other hybrid models include the RETINA model of Masuad et al. [111]; the multichannel convolutional bidirectional gated recurrent unit (MCBiGRU) [112]; attention-LSTM of Wang and Ding [113]; bi-directional GRU layers with CNN of Wiedemann et al. [114], Setyadi et al. [115], and Ziqi et al. [87]; variation dataset of Gambäck et al. [69] based on the Tsironi [116] model; CAT boost of Qureshi and Sabih [117]; pre-trained word embeddings and metadata [118,119,120]; ensemble RNN of Pitsilis et al. [108] and [121]; LSTM + Random Embedding + Boosted Decision Trees (GBDTs) of Joulin et al. [109] and Miok et al. [122]; MANDOLA of Paschalides et al. [110]; MCD + LSTM [122], Sachdeva et al. [123], Sajjad et al. [124], and Liu et al. [125]; the fuzzy rule method [126], Ayo et al. [7]; the KDEHatEval team [121], Ali et al. [127], and J et al. [112]; and Google’s Jigsaw on the Kaggle Model (https://www.kaggle.com/c/jigsaw-toxic-comment-classification-challenge/overview, accessed on 10 February 2022), Wang and Ding [113]; attention-LSTM model based on a modified version of HAN [128] and a BiGRU-capsule model using the dataset of [46], Wiedemann et al. [114]; bi-directional GRU layers with CNN layers, Setyadi et al. [115], Ziqi et al. [87], and Qureshi and Sabih [117].

3.2. Datasets

Datasets have targeted numerous hate speech categories, as shown in Appendix B. However, examples could be unclear in several datasets, such as Waseem’s dataset [27] or hierarchical datasets including the dataset of Basile et al. [58]. Furthermore, datasets are of poor quality because they are not regularly updated when Twitter users use new phrases or abbreviations. In addition, approximately 60% of dataset creators found inter-annotator agreement [1]. Therefore, a useful predictive hate speech detection model requires relevant and non-obsolete datasets. The maturity of datasets is regarded as a one-of-a-kind task for superior quality systems. According to Koco et al. [129], separating annotator groups has a more significant impact on the performance of hate detection systems. They also stated that group consensus impacts the quality of recognition. It has been demonstrated that the identification group of people who publish tweets introduces bias into the dataset, and negative data are challenging to compile and ensure; hence, implicit hate speech is difficult to measure [130]. In addition, many datasets overlap across class labels, as shown by Waseem [50], who found an overlap of 2876 tweets between the Waseem and Hovy dataset.

Our analysis showed that research requires more robust, trustworthy, and large datasets due to the wide applications of hate speech detection. The dataset of Shibly et al. [65] developed a robust and colossal dataset. Those datasets backed by conference or workshop series, such as Hate Speech and Offensive Content Identification in Indo-European Languages (HASOC) [19,38] and SemEval, are probably among the most popular datasets. HASOC is divided into three subtasks: the first focuses on identifying hate speech and offensive language (sub-task A); the second focuses on identifying the type of hate speech (sub-task B); and the third focuses on identifying the hate speech’s target group (or persons). The SemEval Task 5: Multilingual Detection of Hate Speech Against Immigrants and Women in Twitter continues to focus on numerous academics in the SemEval series [83,113]. On the other hand, researchers have paid close attention to SemEval 2019 Task 6 [18]—OffensEval: Identifying and Categorizing Offensive Language on Social Media. The Offensive Language Identification Dataset (OLID) [101], which contains over 14,000 English tweets, is the most recent (similar tasks as in HASOC). The HASOC 2020 track dataset [19] includes 3708 English tweet samples, however, it is considered a substantial and competitive dataset. Nearby random (reported F1-score of 0.52) had the best performance, whereas fine-grained multilabel categorization had the worst performance (0.26 F1-measure). ElSherief et al. [131] found 25,278 hate instigators and 22,857 target users in their study. Note that datasets that were not built for Twitter (e.g., [132,133]) were not evaluated, and datasets that were not built for English were discarded (e.g., [134]). According to our reasoning, such datasets are not appropriate due to the variety of beneficial qualities. The list of reviewed datasets is shown in Appendix B.

4. Discussion

The analysis of collected papers revealed that defining hate speech is still an issue, as previously discussed [7]. According to Culpeper [135], incitement, face, and impoliteness are all related to hate speech, but they are distinct; the author stated that mapping between the concepts was complicated due to the Oxford English Corpus containing multiple genres during the period of 2000–2006, with varying frequencies of hate and impoliteness. To comprehend white supremacist heteropatriarchy, intersectionality between race, sex, place, and gender has been found to represent a framework of hate speech-related notions on social media [13]. Racist and homophobic tweets were shown to be more hateful than offensive sexist ones [25]. Wasim and Hovy [28] described offensive tweets using 11 criteria, including racial slurs, attacks on minorities, supporting hate speech, misrepresenting the truth, and defending xenophobia or sexism. As a result, hate speech has many meanings, including offensive, abusive, harassing, aggressive, cyberbullying, destructive, and hate speech [136]. However, offensive comments, such as hate speech, are among the most researched NLP topics [129]. According to MacAvaney [10], there is no standard hate speech definition. Nevertheless, according to Schmidt and Wiegand [11], hate speech is a broad phrase for various disparaging content produced by humans.

Hate speech detection is a complex process, partly due to the availability of the datasets and the task of developing a machine learning model with rigid performance.

Table 1. Themes of papers.

Category of Papers	Meaning	Papers
Small datasets	A dataset is considered small if it has less than the initial dataset of Waseem, which was 16 K tweets.	[33,41,49,55,76,81,115,115]
Binary classification	A model is considered a binary classification model if it presents a work that classifies hate speech into two or three classes.	[26,28,28,31,37,37,41,41,46,51,59,65,65,67,69,69,78,80,80,90,93,94,108,109,114,115,115,113,61,121,88,123,114,126,110,113,87]
Low performance	A low-performance classifier is considered as that which reports a binary classification below 0.6.	[22,22,22,28,30,41,44,46,47,49,55,59,59,59,78,82,121,138]
Topic-wise detection	A study that uses topic-wise categorization instead of classification.	[41,46,71,76,76,117,117,139]
Lexicon-based dictionary generalization	A low-performance classifier is considered as those which report a binary classification below 0.6.	[24,102,104,106,122,140]
Semi-supervised, clustering, rule-based	A study that uses semi-supervised learning or rule-based methods instead of classification to solve the issue of hate detection.	[7,113,125,141]
Specialized (one dimension of hate speech)	Focused on a specific category of hate speech or dimension of hate speech.	Sarcasm [116] Racism [66,70] Sexism [25,66] General sexist tweets hide a sentiment of hate or misogynistic attitude [34] Detecting profiles [53,54] UK only [40] Retweeting [111] Hate intensity [141] Multi-mixed languages [56] Multi-hierarchical classification [142] Hate speech against immigrants [46] Comments and large text [74,112,143,144]

divides the essential themes of the papers under consideration into seven categories. As shown in Figure 5, most publications use machine learning models to categorize information into binary classifications with a maximum of three class labels. The HASOC and CrowdFlower datasets are uneven in terms of the positive class, with more offensive tweets in CrowdFlower and more classed as ‘neither’ in HASOC, as stated in many research studies. Furthermore, detecting abusive language is problematic for various reasons, including word obfuscation, difficulties tracing racist and slurs, and professionally written and abusive language that crosses sentence boundaries. Several projects (such as the SemEval and HASOC tasks) require more robust, trustworthy, and massive datasets due to the wide applications of hate speech detection. However, small datasets are insufficient for generalizing conclusions or capturing features. For example, some research merely employed 200 tweets, generating good results but raising concerns about the reported performance’s generality. Therefore, the size of the dataset does not guarantee diversity [137]. Appendix B shows the most relevant papers and their limitations. It was not practical to list all 138 papers; therefore, we selected 23 papers from the themes discussed in Section 3.1. We tried to present different methods that are published in different datasets.

4.1. Challenges of Machine Learning Models

The literature typically utilizes the following algorithms based on the described standard machine learning hate detection methods: support vector machines (SVM); Naive Bayes (NB); Logistic Regression (LR); Decision Trees (DTs); and K-Nearest Neighbor (KNN). While the approaches’ performance varies depending on the dataset, the LR consistently outperforms the others. Most algorithms rely on n-grams, POS tags, and feelings to detect hate speech, while some integrate lexicons as a bonus feature. However, based on the varied datasets, these techniques produce contradictory conclusions. The fact that these methods have been used for binary classification jobs is one of the major concerns. As a result, without considering the word’s context and reviewing vast and robust datasets, their performance may, in most situations, be in direct opposition to deep learning models.

The most promising algorithms in deep learning methods, according to research, are CNN and LSTM. For example, Anand and Eswari [139] explored LSTM and CNN with and without GloVe embeddings. Using the same dataset as CNN, they reported an accuracy of above 97% (GloVe and CNN = 97.27). However, according to the research, the approaches have varying performance levels, and the majority of them focus on a binary hate detection problem (hate/not hate) with multilabel classifications only occasionally being explored. One could be due to the lack of a dataset, while the other could be due to the difficulties of detecting hate speech.

Nonetheless, CNN and LSTM have shown to be effective in several trials. Due to the modest size of the datasets, pre-trained BERT models produce varied outcomes. As a result, the problem with deep machine learning models is similar to past problems with classic features approaches and dataset size. According to the literature, the performance of most hybrid approaches has improved by at least 5% compared to previous detection methods.

Consequently, hybrid machine learning models are dominant in the literature and show promising results; however, without large datasets, generalization remains impossible.

4.2. Challenges of Datasets

Comparing hate speech detection methods is difficult due to the variety of features and datasets. With the imbalanced class distribution, hate speech lacks a discriminative or unique collection of features [140]. Depending on the notability and features or algorithms used, traditional approaches (such as SVM and NB) can outperform deep learning models. For example, the SVM method in [31] outperforms the Hierarchical Attention Network (HAN) [128] and CNN utilizing Glove word embedding [31]. Nugroho et al. [145] showed that RF outperforms a neural network model by approximately 10% (0.72). ElSherief et al. [131] found 25,278 hate instigators and 22,857 target users in their study. They discovered that agreeableness, openness, emotional range, conscientiousness, and extraversion all play a role in detecting the profiles of instigators. They also discovered that hate instigators target more visible users with common personality features, such as anger, sadness, and immoderation. Shallow lexical features [25], dictionaries [146], sentiment analysis [147], linguistic characteristics [148], knowledge-based features [149], and meta-information [21] were described in the literature as features connected to tweets.

Consequently, the major issue of datasets is how to prepare a large dataset without being biased during the manually performed annotation process. The second challenge is related to automated methods, which depend on keywords that might have conflicting definitions, as previously illustrated in Section 4.

4.3. Challenges of Feature Sets

Despite significant efforts to develop identifying traits for hate speech, the subject remains open due to differing definitions of hate speech and new online hate vocabulary. There was no agreement in the literature on what constitutes hate speech [150]. For example, Davidson et al. [25] differentiated between profanity, insults, and hate speech, whereas Warner and Hirschberg [39] defined abuse as based on a person’s intrinsic qualities (e.g., ethnicity, sexual orientation, or gender). As a result, it is impossible to say whether the objectionable text is hateful or not [8]. Therefore, hate speech is a culturally dependent term with a specific definition [41]. Furthermore, hate speech may be interpreted in various ways depending on the level of democracy or policies enacted in each country, which may or may not correspond to international standards to some extent.

Feature detection [151,152] and identifying hate speech at the level of each class label, rather than in terms of a binary classification (hate and no hate), are discussed in this study. A small number of studies have reported the performance of hate speech and related domains in relation to class labels [140]. Furthermore, datasets are frequently uneven, with hateful class occurrences far outnumbering non-hate textual examples, making micro-F1-scores appear high due to the majority class. Many machine learning approaches work well on specialized datasets, but they are incapable of generalizing new hate speech content [82].

Consequently, the issue is finding a set of features that is sufficient to generalize to new datasets. Additionally, the feature set should be minimal, providing high-performance models. However, to the best of our knowledge, little research has discussed a generic model that applies to any domain of interest or any dataset.

4.4. Future Research Directions

Based on the discussed literature, the following are future research directions in the domain of textual hate speech detection:

• There is a critical dearth of reporting in the literature on the optimal set of features for hate speech detection that can be applied to both classical and deep learning models. Therefore, extensive research is needed to develop features that work well with diverse datasets with multifaceted hate speech concepts. A successful model should also have features that can be applied to new datasets and previously unseen tweets. A direction could be research [45,153] in which more features are added to develop additional features.
• Aside from the basic hate/no hate categorization for traditional and deep learning models, the literature lacks a detailed investigation of fine-grained hate speech detection at the label level. According to the studies gathered, there is still a gap in creating a model that successfully performs the multi-classification of hate speech, has acceptable performance, and can be generalized across settings. A starting point could be using the models of [81], where several classes were adopted.
• There are no recommendations in the literature to ensure that hate speech detection methods are adequately compared across different datasets. Therefore, a new methodology for dataset comparison is needed so that datasets can be rigorously compared.

5. Conclusions

New datasets of hate speech from different regions with various topics of hate speech and offensive content are constantly being developed. However, many datasets are small in size whilst others lack reliability due to how the data were collected and annotated. Additionally, many datasets are small or sparse, lacking linguistic variety. Above all, the language of the content and region where the data were collected from social media make the comparison between various hate speech detection models difficult. One of the significant challenges in hate speech detection is the architecture of the machine learning model and the lack of consensus on hate speech definition. It was reported that creating large and varied hate or abusive datasets that minimize potential bias is laborious and requires specialized experts. The critical review provided evidence that the research community should enhance the currently developed datasets and constantly update them. Although many machine learning models are developed, feature set selection should be fine-grained at the label level, as well as in the generic hate/no hate classification by developing the best set of features for hate speech detection, which is applicable to traditional and deep learning models. The literature lacks guidelines that assure a proper comparison between the methods over different datasets.