Hate Speech Detection Using NLP and Machine Learning

The aim of this paper is to review machine learning (ML) algorithms and techniques for hate speech detection in social media (SM). Hate speech problem is normally model as a text classification task. In this study, we examined the basic baseline components of hate speech classification using ML algorithms. There are five basic baseline components – data collection and exploration, feature extraction, dimensionality reduction, classifier selection and training, and model evaluation, were reviewed. There have been improvements in ML algorithms that were employed for hate speech detection over time. New datasets and different performance metrics have been proposed in the literature. To keep the researchers informed regarding these trends in the automatic detection of hate speech, it calls for a comprehensive and an updated state-of-the-art. The contributions of this study are three-fold. First to equip the readers with the necessary information on the critical steps involved in hate speech detection using ML algorithms. Secondly, the weaknesses and strengths of each method is critically evaluated to guide researchers in the algorithm choice dilemma. Lastly, some research gaps and open challenges were identified. The different variants of ML techniques were reviewed which include classical ML, ensemble approach and deep learning methods. Researchers and professionals alike will benefit immensely from this study.

In recent times, the Internet and social media are very well-known and popular among people. Usage of social media is increased exponentially during the last few years globally, and it allows people to engage with one another and share ideas, thoughts, opinions, etc. Every day, massive amounts of data are disseminated at breakneck speed via social media platforms, reaching a massive audience. Furthermore, the ability to write anonymous posts and comments makes expressing and spreading hate speech even easier. To improve the users’ experience, social media sites are attempting to remove hateful remarks. In this research work, the main focus is on developing automated hate speech and offensive language detection models. Started with traditional hate speech and offensive language identification approaches, than reaches to advanced hate speech recognition methods for social media. This brings a need for integrated datasets and the hate speech prediction method. This paper describes the study on hate speech and offensive content identification in English language by using the various approaches based on machine learning algorithms (Support vector machine, decision tree, and so on) and NLP, along with the features used for the classification problem. The models were tested on the HASOC (2021) datasets and concluded that the ensemble model perform better than other algorithms with the test dataset for different tasks. Results and analysis part of this paper offers researchers a comprehensive picture of approaches.

Hate speech towards a group or an individual based on their perceived identity, such as ethnicity, religion, or nationality, is widely and rapidly spreading on social media platforms. This causes harmful impacts on users of these platforms and the quality of online shared content. Fortunately, researchers have developed different machine learning algorithms to automatically detect hate speech on social media platforms. However, most of these algorithms focus on the detection of hate speech that appears in English. There is a lack of studies on the detection of hate speech in Arabic due to the language’s complex nature. This paper aims to address this issue by proposing an effective approach for detecting Arabic hate speech on social media platforms, namely Twitter. Therefore, this paper introduces the Arabic BERT-Mini Model (ABMM) to identify hate speech on social media. More specifically, the bidirectional encoder representations from transformers (BERT) model was employed to analyze data collected from Twitter and classify the results into three categories: normal, abuse, and hate speech. In order to evaluate our model and state-of-the-art approaches, we conducted a series of experiments on Twitter data. In comparison with previous works on Arabic hate-speech detection, the ABMM model shows very promising results with an accuracy score of 0.986 compared to the other models.

The CASE@EACL2024 Shared Task addresses Climate Activism online through three subtasks that focus on hate speech detection (Subtask A), hate speech target classification (Subtask B), and stance detection (Subtask C) respectively.Our contribution examines the effect of finetuning on external data for each of these subtasks.For the two subtasks that focus on hate speech, we augment the training data with the OLID (Zampieri et al., 2019a) dataset, whereas for the stance subtask we harness the SemEval-2016 Stance dataset (Mohammad et al., 2016b).We fine-tune RoBERTa and DeBERTa models for each of the subtasks, with and without external training data.For the hate speech detection and stance detection subtasks, our RoBERTa models came up third and first on the leaderboard, respectively.While the use of external data was not relevant on those tasks, we found that it greatly improved the performance on the hate speech target categorization.

Hate speech is a form of negative communication intended to harm people and communities. Hate speech is quite common in the real world, and it has reached alarming proportions on social media as well. These days our lives have become increasingly reliant on social media platforms, such as Facebook. This is due to the rapid advancement of technology and communication. In Bangladesh, the number of people using social media platforms is also rapidly increasing. In English, detecting hate speech on social media is a difficult task. Comparatively, Bengali is a complicated language with few datasets available. As a result, detection of Bengali hate speech becomes even more challenging. In this paper, we present a new dataset of 10,133 user comments. We have collected them from the comment section of various public Facebook pages. We explore the performance of various machine learning and deep learning models in detecting hate speech. Bengali pre-trained word embeddings from fastText are used to train the models. We are especially interested in Convolutional Neural Network (CNN). To our knowledge it was never used for hate speech detection in binary classification. Another goal of this research is to create a new and large dataset, which will facilitate further research of Bengali Hate Speech Detection. All machine learning and deep learning models performed very well from our experiments. But, Support Vector Machine (SVM) is the one that performed the best among them.

Technological developments over the past few decades have changed the way people communicate, with platforms like social media and blogs becoming vital channels for international conversation. Even though hate speech is vigorously suppressed on social media, it is still a concern that needs to be constantly recognized and observed. The Arabic language poses particular difficulties in the detection of hate speech, despite the considerable efforts made in this area for English-language social media content. Arabic calls for particular consideration when it comes to hate speech detection because of its many dialects and linguistic nuances. Another degree of complication is added by the widespread practice of "code-mixing," in which users merge various languages smoothly. Recognizing this research vacuum, the study aims to close it by examining how well machine learning models containing variation features can detect hate speech, especially when it comes to Arabic tweets featuring code-mixing. Therefore, the objective of this study is to assess and compare the effectiveness of different features and machine learning models for hate speech detection on Arabic hate speech and code-mixing hate speech datasets. To achieve the objectives, the methodology used includes data collection, data pre-processing, feature extraction, the construction of classification models, and the evaluation of the constructed classification models. The findings from the analysis revealed that the TF-IDF feature, when employed with the SGD model, attained the highest accuracy, reaching 98.21%. Subsequently, these results were contrasted with outcomes from three existing studies, and the proposed method outperformed them, underscoring the significance of the proposed method. Consequently, our study carries practical implications and serves as a foundational exploration in the realm of automated hate speech detection in text.

Racist and ethnic violence, fabricated persecution, and some form of intimidation are all risks associated with hate speech, which is a concern with natural language processing. Given the sensitivity of hate speech in our society, it is essential to classify speeches into hate and non-hate categories in real time to minimize its risks. The main objective of this work is to investigate selected supervised machine learning algorithm model for the classification of hate speech on social media. The term frequency-inverse document frequency (TF-IDF) and bag of words (BOW) models were used by the model to extract features. Porter’s stemming model and WordNet for lemmatization are used during the preprocessing step. The datasets were trained using logistic regression, naive Bayes, and random forest, and logistic regression was also utilized to create the classifier. For training purpose, 80% of the datasets was used to train the model and 20% was used for testing the model. Results obtained from the application of Logistic Regression algorithm revealed 98% accuracy and 98% F1-score. These scores indicate high accuracy in hate speech detection and classification.

Online hate has emerged as a rapidly growing issue worldwide, often stemming from differences in opinion. It is crucial to use appropriate language and words on social media platforms, as inappropriate communication can negatively impact others. Consequently, detecting hate speech is of significant importance. While manual methods are commonly employed to identify hate and offensive content on social media, they are time-consuming, labor-intensive, and prone to errors. Therefore, AI-based approaches are increasingly being adopted for the effective classification of hate and offensive speech. The proposed model incorporates various text preprocessing techniques, such as removing extraneous elements like URLs, emojis, and blank spaces. Following preprocessing, tokenization is applied to break down the text into smaller components or tokens. The tokenization technique utilized in this study is TF-IDF (Term Frequency–Inverse Document Frequency). After tokenization, the model performs the classification of hate and offensive speech using the proposed BiLSTM-based SM-CJ (Scalable Multi-Channel Joint) framework. The BiLSTM-based SM-CJ model is particularly effective in detecting hate, offensive, and neutral tweets due to its ability to capture both forward and backward contexts within a given text. Detecting hate speech requires a comprehensive understanding of the text and the identification of patterns spanning across multiple words or phrases. To achieve this, the LSTM component of the BiLSTM model is designed to capture long-term dependencies by utilizing information from earlier parts of the text. The proposed SM-CJ framework further aligns the input sequence lengths fetched from the input layer, enabling the model to focus on specific segments of the input sequence that are most relevant for hate speech detection. This approach allows the model to accurately capture derogatory language, and subtle nuances present in hate speech. Finally, the performance of the proposed framework is evaluated using various metrics, including accuracy, recall, F1-score, and precision. The results are compared with state-of-the-art approaches, demonstrating the effectiveness of the proposed model. Doi: 10.28991/ESJ-2025-09-01-03 Full Text: PDF

Detecting hate speech in social media is challenging due to its rarity, high-dimensional complexity, and implicit expression via sarcasm or spelling variations, rendering linear models ineffective. In this study, the SVM (Support Vector Machine) algorithm is used to map text features from low-dimensional to high-dimensional space using kernel function techniques to meet complex nonlinear classification challenges. By maximizing the category interval to locate the optimal hyperplane and combining nuclear techniques to implicitly adjust the data distribution, the classification accuracy of hate speech detection is significantly improved. Data collection leverages social media APIs (Application Programming Interface) and customized crawlers with OAuth2.0 authentication and keyword filtering, ensuring relevance. Regular expressions validate data integrity, followed by preprocessing steps such as denoising, stop-word removal, and spelling correction. Word embeddings are generated using Word2Vec’s Skip-gram model, combined with TF-IDF (Term Frequency–Inverse Document Frequency) weighting to capture contextual semantics. A multi-level feature extraction framework integrates sentiment analysis via lexicon-based methods and BERT for advanced sentiment recognition. Experimental evaluations on two datasets demonstrate the SVM model’s effectiveness, achieving accuracies of 90.42% and 92.84%, recall rates of 88.06% and 90.79%, and average inference times of 3.71 ms and 2.96 ms. These results highlight the model’s ability to detect implicit hate speech accurately and efficiently, supporting real-time monitoring. This research contributes to creating a safer online environment by advancing hate speech detection methodologies.

Hate speech has long been a prevalent issue offline, but with the rise of the internet and social media, its spread has accelerated. The anonymity afforded by these platforms enables individuals to engage in hate speech often without facing substantial consequences. As technology continues to evolve, new opportunities arise to tackle this problem, particularly through the use of natural language processing (NLP). NLP technology can help automate processes that have traditionally been done manually, such as flagging hate speech in online content. Yet, many issues remain unresolved before we can achieve efficient hate speech detection systems. These include foundational challenges, such as defining hate speech, as well as issues that arise at the end of an NLP pipeline, such as evaluating whether a model can generalize. Additionally, the issue of bias in models poses a significant challenge, as biased training data can lead to inaccurate or unfair results. In this thesis, I will focus on addressing these issues. First, I examine the definitions of hate speech and related concepts like toxicity and abusive language, and their impact on re-annotated datasets. I then compare the original and re-annotated labels in terms of robustness and generalization using a BERT-based classifier. Next, I explore the use of hate speech legislation from three countries for annotation, expanding the task of hate speech detection to prosecutable hate speech detection. The results show even law interpretation can be subjective, which has a consequent effect on model training and evaluation. To address this issue, I introduce a semantic componential analysis of hate speech definitions, leading to the creation of the HateDefCon corpus. Moreover, I present a pipeline for generating parallel multilingual hate speech corpora and discuss the associated challenges. The thesis concludes with an examination of textual biases based on the psycholinguistic aspects of harmful language.

Social media sites facilitate users’ discussions, expression of opinions, sharing of information and news, and promotion of ideas and products, thus rapidly increasing the volume of hate speech and offensive content on online platforms. Consequently, hate speech and offensive content have turned out to be a widespread issue that negatively affects both individuals and society and needs to be controlled through detection and removal. This paper aims to provide a further understanding of the meaning of hate speech and offensive language and to provide a comprehensive discussion of the various techniques to detect hate speech and offensive language. A systematic literature review (SLR) of 90 research papers published between 2018 and 2024 was conducted to discover gaps within the literature. This review revealed challenges and possibilities for further development and improvement of previous findings. The results show that most of these works classified hate speech and offensive language using techniques from machine learning (ML) and deep learning (DL), and the most common performance metrics were Accuracy, Precision, Recall, and F‐measure. The benchmark datasets are also described. Twitter was the most commonly utilized social network for obtaining datasets, while Facebook is sometimes used. Moreover, the findings of this review offer insight into research trends in Arabic hate speech and offensive language, as well as new research directions. The most interesting finding is that until now most Social Media Network Developers have not included autodetection of hate speech or offensive language plugins. Finally, this study presents a guideline for choosing the best strategies and techniques to detect and predict Arabic offensive language and hate speech.

Over the last two decades, social media platforms have grown dramatically. Twitter and Facebook are the two most popular social media platforms, with millions of active users posting billions of messages daily. These platforms allow users to have freedom of expression. However, some users exploit this facility by disseminating hate speeches. Manual detection and censorship of such hate speeches are impractical; thus, an automatic detection mechanism is required to detect and counter hate speeches in a real-time environment. Most research in hate speech detection has been carried out in the English language. Still, minimal work has been explored in other languages, mainly Urdu written in Roman Urdu script. A few research have attempted machine learning, and deep learning models for Roman Urdu hate speech detection; however, due to a scarcity of Roman Urdu resources, and a large corpus with defined annotation rules, a robust hate speech detection model is still required. With this motivation, this study contributes in the following manner: we developed annotation guidelines for Roman Urdu Hate Speech. Second, we constructed a new Roman Urdu Hate Speech Dataset (RU-HSD-30K) that was annotated by a team of experts using the annotation rules. To the best of our knowledge, the Bi-LSTM model with an attention layer for Roman-Urdu Hate Speech Detection has not been explored. Therefore, we developed a context-aware Roman Urdu Hate Speech detection model based on Bi-LSTM with an attention layer and used custom word2vec for word embeddings. Finally, we examined the effect of lexical normalization of Roman Urdu words on the performance of the proposed model. Different traditional as well as deep learning models, including LSTM and CNN models, were used as baseline models. The performance of the models was assessed in terms of evaluation matrices like accuracy, precision, recall, and F1-score. The generalization of each model is also evaluated on a cross-domain dataset. Experimental results revealed that Bi-LSTM with attention outperformed the traditional machine learning models and other deep learning models with an accuracy score of 0.875 and an F-Score of 0.885. In addition, the results demonstrated that our suggested model (Bi-LSTM with Attention Layer) is more general than previous models when applied to unseen data. The results confirmed that lexical normalization of Roman Urdu words enhanced the performance of the suggested model.

This article investigates the identification of hate speech on social media using machine learning and deep learning techniques. The research uses metrics such as F-measure, AUC-ROC, precision, accuracy, and recall assessing the effectiveness of various tactics. The findings indicate that deep learning models, particularly the bidirectional long short-term memory (BiLSTM) architecture, consistently outperform other methods in categorization tasks. The research emphasizes the importance of sophisticated neural network designs in identifying the intricacies of hostile and offensive content online. The study offers insights for promoting early identification and prevention of cyberbullying, improving secure and inclusive online environments. Future research may explore real-time detection systems, hybrid approaches, or the integration of complementary components to enhance and improve innovative technology in tackling this significant social issue. A sample tweet was annotated by specialists who categorize tweets as hate speech, offensive language, or neutral. The researchers applied shallow learning methodologies and integrated word embeddings like Word2Vec and GloVe to enhance the efficacy of deep learning models. The results indicate that BiLSTM surpasses shallow learning methods in detecting hate speech on Twitter, highlighting the efficacy of deep learning approaches in recognizing and tracking hate speech on social media platforms. When comparing different deep learning and machine learning models on different datasets, the results reveal that deep learning techniques are usually more effective. A reasonably high level of accuracy is achieved by KNN and SVM among classical algorithms, whereas Naïve Bayes performs the poorest. While deep learning approaches provide better results, tree-based models such as Random Forest and Decision Trees offer more consistent accuracy. Models based on neural networks, such as LSTM, CNN, and BI-LSTM, perform well, with LSTM-based methods excelling in particular. The most successful strategy for classification problems is the model presented, which obtains the greatest accuracy, precision, recall, F1-score of 95%. The research aids in the development of advanced tools and methodologies to mitigate hate speech on social media and foster positive online interactions. Future research may investigate alternative deep learning architectures, such as transformers, to enhance hate speech detection efficacy. The advancement of interpretable AI methodologies for identifying hate speech and delivering transparent forecasts might enhance user confidence and facilitate better content moderation decisions.

In a text classification problem, imbalance nature in dataset oftentimes has been disregarded even though it might significantly have an impact on the results of the classification model performance. This issue also occurred in the hate speech detection where most collected datasets are highly unbalanced. Among state-of-the-art methods that deal with classifying disparity data, the sampling-based technique is the most effective approach in classifying an imbalanced data. In this paper, four resampling methods include Random Oversampling (ROS), Synthetic Minority Technique (SMOTE), Adaptive Synthetic (ADASYN) and Random Undersampling (RUS) are used as an answer to the inequality of class distribution in a hate speech dataset. With three basic machine learning classifiers i.e. Support Vector Machine, Logistic Regression and Naïve Bayes, the evaluation results show that the oversampling approach improves the accuracy and the overall performance of three classifiers. Among all resampling techniques and machine learning algorithms, Logistic Regression enforced by ROS performed the best with an overall accuracy of 91 percent and F1-Score of 0.95.

In these contemporary times, social media is omnipresent and most people adhere to at least one of these digital platforms. Social entertainment generates an enormous amount of data and this is an unparalleled opportunity for data scientists and linguistic experts. These factors have renewed the interest in Natural Language Processing techniques and as such, there is a continuous increase in the number of publications that deal with the topic of Tweet classification using machine learning models. In this paper, experiments performed by the TweetEval team from the University of Cardiff have been studied and expanded upon. These tasks include emotion detection, offensive language identification and hate speech detection. The decision was made to focus on these specific classification tasks as they directly relate to unsought behaviours such as online harassment. This research endeavour involved building and testing a transformer-based language model which is capable of matching the performance of TweetEval. The aim of this study is therefore to identify common limitations to such models and how these can be circumvented to effectively combat phenomenon such as cyberbullying and online abuse using machine learning. From the results that were obtained, the developed BERT model performed comparatively well to other similar algorithms for all tasks as the obtained results were an F1-Score of 0.51, 0.76 and 0.80 for hate speech, emotion detection and offensive language respectively.