Hate Speech Detection Research Articles

MBERT-GRU multilingual deep learning framework for hate speech detection in social media

One major issue plaguing online social media is hate speech, a complex phenomenon whose identification and target categorization have been studied by the natural language processing community. In recent years, notable studies have been made towards hate speech detection using various mechanisms varying from traditional machine learning to complex deep neural network models. However, these studies mainly focus on high-resource English language. The multilingual societies such as the Indian subcontinent: English, Hindi and Hindi-English code-mixed languages are widespread and convenient for the users. The research works studying hate speech detection in these languages are still very limited. To fill this gap, we propose an mBERT-GRU framework comprising of multilingual BERT embedding and bidirectional GRU layers to learn the cumulative features for hate speech detection and its target categorization. We evaluated our work on three datasets HASOC-2019, HS and HEOT to prove the competitive performance. Our results show that the proposed framework outperformed monolingual and state-of-the-art methods on English, Hindi and Hindi-English code-mixed datasets with Macro-F1 measure values of 0.87, 0.83 and 0.77, respectively.

Deep learning models used for accuracy and loss detection on twitter dataset by setting tuning parameters

Hate speech on social media post is running now a days. Social media like YouTube, Twitter, and Facebook etc. are responsible for hated speech. Hated speech spreads through digital media, causing individuals to get confused and adopt prejudiced viewpoints. To limit the negative effects of disinformation on the digital platform, it is critical to detect it. Now a days, lots of digital platforms are available. Hate speech detection in dataset is very difficult. As a result, the Twitter dataset is of the size of 25296 is presented in this work. Many deep learning techniques are applied on Twitter dataset. The Google Colab tool is used to scrape dataset material. Different deep learning approaches are utilized to boost the accuracy of the hated speech dataset. For training and validation accuracy and loss some models are used on Twitter dataset like Bi-directional Long Short Term Memory with Glove, Bi-LSTM, and Embedding from Language Model (Elmo) with deep learning, Convolutional Neural Network (CNN), Long Short Term Memory with Glove and LSTM. The performance of the proposed tweet dataset is evaluated using a variety of deep learning classifiers on text dataset. The planned deep learning techniques produced good results on tweet dataset. LSTM with Glove gave the highest accuracy 0.89 and minimum loss 0.19 on tweet dataset. So when compare our model on same dataset that was used earlier then we get highest accuracy and minimum loss.

OLID-BR: offensive language identification dataset for Brazilian Portuguese

Social media has revolutionized the manner in which our society is interconnected. While this extensive connectivity offers numerous benefits, it is also accompanied by significant drawbacks, particularly in terms of the proliferation of fake news and the vast dissemination of hate speech. Identifying offensive comments is a critical task for ensuring the safety of users, which is why industry and academia have been working on developing solutions to this problem. Prior research on hate speech detection has predominantly focused on the English language, with few studies devoted to other languages such as Portuguese. This paper introduces the Offensive Language Identification Dataset for Brazilian Portuguese (OLID-BR), a high-quality NLP dataset for offensive language detection, which we make publicly available. The dataset contains 6,354 (extendable to 13,538) comments labeled using a fine-grained three-layer annotation schema compatible with datasets in other languages, which allows the training of multilingual/cross-lingual models. The five NLP tasks available in OLID-BR allow the detection of offensive comments, the classification of the types of offenses such as racism, LGBTQphobia, sexism, xenophobia, and so on, the identification of the type and the target of offensive comments, and the extraction of toxic spans of offensive comments. All those tasks can enhance the capabilities of content moderation systems by providing deep contextual analysis or highlighting the spans that make a text toxic. We further experiment with and evaluate the dataset using state-of-the-art BERT-based and NER models, which demonstrates the usefulness of OLID-BR for the development of toxicity detection systems for Portuguese texts.

A web framework for information aggregation and management of multilingual hate speech

Social media platforms have led to the creation of a vast amount of information produced by users and published publicly, facilitating participation in the public sphere, but also giving the opportunity for certain users to publish hateful content. This content mainly involves offensive/discriminative speech towards social groups or individuals (based on racial, religious, gender or other characteristics) and could possibly lead into subsequent hate actions/crimes due to persistent escalation. Content management and moderation in big data volumes can no longer be supported manually. In the current research, a web framework is presented and evaluated for the collection, analysis, and aggregation of multilingual textual content from various online sources. The framework is designed to address the needs of human users, journalists, academics, and the public to collect and analyze content from social media and the web in Spanish, Italian, Greek, and English, without prior training or a background in Computer Science. The backend functionality provides content collection and monitoring, semantic analysis including hate speech detection and sentiment analysis using machine learning models and rule-based algorithms, storing, querying, and retrieving such content along with the relevant metadata in a database. This functionality is assessed through a graphic user interface that is accessed using a web browser. An evaluation procedure was held through online questionnaires, including journalists and students, proving the feasibility of the use of the proposed framework by non-experts for the defined use-case scenarios.

Hate Speech Detection Using Machine Learning

Abstract: Twitter's central goal is to enable everybody to make and share thoughts and data, and to communicate their suppositions and convictions without boundaries. Twitter's job is to serve the public discussion, which requires portrayal of a different scope of points of view. Yet, it does not advance viciousness against or straightforwardly assault or undermine others based on race, nationality, public cause, rank, sexual direction, age, inability, or genuine illness. Hate Speech can hurt a person or a community. So, it is not appropriate to use hate speech. Now, due to increase in social media usage, hate speech is very commonly used on these platforms. So, it is not possible to identify hate speeches manually. So, it is essnetial to develop an automated hate speech detection model and this resaech work shows different approaches of Natural Language Processing for classification of Hate Speech through Machine Learning Algorithms.

Detecting Dravidian Offensive Posts in MIoT: A Hybrid Deep Learning Framework

Hate speech and Offensive Posts (OP) detection on Smart Multimedia Internet of Things (MIoT) have been an active issue for researchers. MIoT media texts in non-native English-speaking countries are often code-mixed or script mixed/switched. This paper proposes an ensemble-based Deep Learning (DL) framework comprised of a Convolutional Neural Network (CNN) and a Dense Neural Network (DNN) for identifying hate and OP in Malayalam Code-Mixed (MCM), Tamil Code-Mixed (TCM), and Malayalam Script-Mixed (MSM) MIoT media postings. Word-level and character-level features are utilized in the convolutional neural network. In contrast, the dense neural network uses character-level Term Frequency-Inverse Document Frequency (TF-IDF) features. The inclusion of character-level features in the proposed ensemble framework resulted in state-of-the-art performance for TCM and MCM datasets, with weighted F 1 -score of 0.91 and 0.78, respectively, and comparable performance for MSM posts, with a weighted F 1 -score of 0.95.

Roman Urdu Hate Speech Detection Using Transformer-Based Model for Cyber Security Applications.

Social media applications, such as Twitter and Facebook, allow users to communicate and share their thoughts, status updates, opinions, photographs, and videos around the globe. Unfortunately, some people utilize these platforms to disseminate hate speech and abusive language. The growth of hate speech may result in hate crimes, cyber violence, and substantial harm to cyberspace, physical security, and social safety. As a result, hate speech detection is a critical issue for both cyberspace and physical society, necessitating the development of a robust application capable of detecting and combating it in real-time. Hate speech detection is a context-dependent problem that requires context-aware mechanisms for resolution. In this study, we employed a transformer-based model for Roman Urdu hate speech classification due to its ability to capture the text context. In addition, we developed the first Roman Urdu pre-trained BERT model, which we named BERT-RU. For this purpose, we exploited the capabilities of BERT by training it from scratch on the largest Roman Urdu dataset consisting of 173,714 text messages. Traditional and deep learning models were used as baseline models, including LSTM, BiLSTM, BiLSTM + Attention Layer, and CNN. We also investigated the concept of transfer learning by using pre-trained BERT embeddings in conjunction with deep learning models. The performance of each model was evaluated in terms of accuracy, precision, recall, and F-measure. The generalization of each model was evaluated on a cross-domain dataset. The experimental results revealed that the transformer-based model, when directly applied to the classification task of the Roman Urdu hate speech, outperformed traditional machine learning, deep learning models, and pre-trained transformer-based models in terms of accuracy, precision, recall, and F-measure, with scores of 96.70%, 97.25%, 96.74%, and 97.89%, respectively. In addition, the transformer-based model exhibited superior generalization on a cross-domain dataset.

Exploring Automatic Hate Speech Detection on Social Media: A Focus on Content-Based Analysis

Hate speech is a challenging problem, and its dissemination can cause potential harm to individuals and society by creating a sense of general unwelcoming to the marginalized groups, which usually are targeted. Therefore, it is essential to understand this issue and which techniques are useful for automatic detection. This paper presents a survey on automatic hate speech detection on social media, providing a structured overview of theoretical aspects and practical resources. Thus, we review different definitions of the term “hate speech” from social network platforms and the scientific community. We also present an overview of the methodologies used for hate speech detection, and we describe the main approaches currently explored in this context, including popular features, datasets, and algorithms. Furthermore, we discuss some challenges and opportunities for better solving this issue.

A hybrid approach based on personality traits for hate speech detection in Arabic social media

<p>In recent years, as social media has grown in popularity, people have gained the ability to freely share their views. However, this may lead to users' conflict and hostility, resulting in unattractive online environments. Hate speech relates to using expressions or phrases that are violent, offensive, or insulting to a minority of people. The number of Arab social media users is quickly rising, and this is being followed by an increase in the frequency of cyber hate speech in the area. Therefore, the automated detection of Arabic hate speech has become a major concern for many stakeholders. The intersection of personality learning and hate speech detection is a relatively less studied niche. We suggest a novel approach that is focused on extracting personality trait features and using these features to detect Arabic hate speech. The experimental results show that the proposed approach is superior in terms of the macro-F1 score by achieving 82.3% compared to previous work reported in the literature.</p>

Detecting Hate Speech and Offensive Language using Machine Learning in Published Online Content

Businesses are more concerned than ever about hate speech content as most brand communication and advertising move online. Different organisations may be incharge of their products and services but they do not have complete control over their content posted online via their website and social media channels, they have no control over what online users post or comment about their brand. As a result, it became imperative in our study to develop a model that will identify hate speechand, offensive language and detect cyber offence in online published content using machine learning. This study employed an experimental design to develop a detection model for determining which agile methodologies were preferred as a suitable development methodology. Deep learning and HateSonar was used to detect hate speech and offensive language in posted content. This study used data from Twitter and Facebook to detect hate speech. The text was classified as either hate speech, offensive language, or both. During the reconnaissance phase, the combined data (structured and unstructured) was obtained from kaggle.com. The combined data was stored in the database as raw data. This revealed that hate speech and offensive language exist everywhere in the world, and the trend of the vices is on the rise. Using machine learning, the researchers successfully developed a model for detecting offensive language and hate speech on online social media platforms. The labelling in the model makes it simple to categorise data in a meaningful and readable manner. The study establishes that in fore model to detect hate speech and offensive language on online social media platforms, the data set must be categorised and presented in statistical form after running the model; the count indicates the total number of data sets imported. The mean for each category, as well as the standard deviation and the minimum and maximum number of tweets in each category, are also displayed. The study established that preventing online platform abuse in Zambia requires a comprehensive approach that involves government law, responsible platform policies and practices, as well as individual responsibility and accountability. In accordance with this goal, the research was effective in developing the detection model. To guarantee that the model was completely functional, it was trained on the English dataset before being applied to the local language dataset. This was because of the fact that training deep learning models with local datasets can present a number of challenges, such as limited, biased data, data privacy, resource requirements, and model maintenance. However, the efficacy of these systems varies, and there have been concerns raised about the inherent biases and limitations of automatic moderation techniques. The study recommends that future studies consider other sources of information such as Facebook, WhatsApp, Instagram, and other social media platforms, as well as consider harvesting local data sets for training machines rather than relying on foreign data sets; the local data set can then be used to detect offences targeting Zambian citizens on local platforms.

Advanced Deep Learning Approaches for Hate Speech Detection

At present, the influence of social networks on a global scale is increasing, the number of users is increasing significantly, and massive amounts of information are being created every day. How to automatically and efficiently identify harmful information, especially hate speech, has become an important issue in the governance of the network environment. In order to improve the screening efficiency of information containing hate speech, deep learning is applied in this field. Although the research work around this task has made great progress, there are very few reviews on this task, lack of a comprehensive review of the latest development in recent years and can not provide help for researchers who are interested in this task. Therefore, we give an overview of the deep learning applied in hate speech, introducing new ideas for solving this task in recent years, and propose potential problems in the task and analyze it.

Inculcating Context for Emoji Powered Bengali Hate Speech Detection using Extended Fuzzy SVM and Text Embedding Models

The massive growth of social webs offer opportunities to communicate with diverse languages, unstructured text, informal posts, misspelled contents and emojis. Social media users feel comfortable to express their emotions specially emotions with high intensity (hate speech) in their mother tongue. Hate speech in any form targets groups and individuals that may trigger antisocial activities, hate crimes, and terrorist acts. Bengali social media users use Bengali for posting implicit or indirect hate text. Existing Bengali hate speech detection research considers explicit hate speech detection but in actual hate is expressed more in implicit way. In order to detect both implicit and explicit hate speech from low resource content, social webs need highly efficient automated tools. Researchers applied discriminative learning approaches (i.e. SVM, MLP, CNN) to distinguish hate text with only clear-cut outcomes in detecting direct hate speech. The proposed novel Bengali hate speech detection model considers two parallel approaches: (i) It applies extended fuzzy SVM classifier for class imbalanced dataset (FSVMCIL) and multilingual BERT (mBERT) text embedding model to detect first hate label; (ii) Morphological analysis method to detect implicit and explicit hate content with the hate similarity (HS) scheme for second hate label. Linking both labeling methods, this research extracts contextual Bengali hate speech from informal text. This novel HS method considers Word2Vec word embedding model and Bengali hate lexicon. It also considers emoji to text conversion for efficient contextual analysis. This study also conducts extensive experiments for various categories with the Bengali hate speech dataset. It also evaluates the proposed model performance considering weighted F1 score, precision, recall and accuracy parameters. Results reveal significant improvement in Bengali hate speech detection with 2.35% increase in F1- score and 9.11 % increase in accuracy.

Indonesian Hate Speech Detection Using Bidirectional Long Short-Term Memory (Bi-LSTM)

Abstract 
 Social media is a platform that allows users to express themselves freely including spreading hate speech content. The government has issued the regulation in the UU ITE to handle and prevent hate speech on social media. The research was also conducted using the Bi-LSTM to classify the text into hate speech or not. Another research was purposed to detect hate speech and its categories using Bi-GRU. However, the performance of the model Bi-GRU is still lower than Bi-LSTM with an accuracy of 86.44% and 96.44%. Therefore, this study aims to build a model that can detect hate speech and its categories. The research offers Bi-LSTM as a classification model and IndoBERT as a tokenization model. The dataset used is a public dataset containing 13 thousand tweets. As a result, the best model obtained is using 20 epochs, 192 batch sizes, 1 layer Bi-LSTM with 40 nodes, and applying class weighing in the optimization process. The pre-train model from IndoBERT that is used to support the performance of the model in classifying is "indobenchmark/indobert-large-p2". The performance given by the purposed model is very good with an average accuracy, precision, and recall of 97.66%, 96.50%, and 85.25%.

Finding hate speech with auxiliary emotion detection from self-training multi-label learning perspective

Hate Speech Detection (HSD) aims to identify whether a text contains hate speech content, which often refers to discrimination and is even associated with a hate crime. The mainstream methods jointly train the HSD problem with relevant auxiliary problems, e.g., emotion detection and sentiment analysis, under the paradigm of Multi-Task Learning (MTL). In this paper, we improve HSD by integrating it with emotion detection, since we take inspiration from the potential correlations between hate speech and certain negative emotion states, which have been studied theoretically and empirically. To be specific, we can concatenate their hateful labels and predicted emotion states as pseudo-multiple labels for hate speech samples, formulating a pseudo-Multi-Label Learning (MLL) problem. Beyond the existing MTL-HSD methods, we further incorporate this pseudo-MLL problem and solve it by capturing the correlations between hate speech and negative emotion states, so as to improve the performance of HSD. Based on these ideas, we propose a novel HSD method named the Emotion-correlated Hate Speech DetectOR (EHSor). We conduct extensive experiments to evaluate EHSor, and the results show that it can consistently outperform the existing HSD methods across benchmark datasets.

HateCircle and Unsupervised Hate Speech Detection Incorporating Emotion and Contextual Semantics

The explosive growth of social media has fueled an extensive increase in online freedom of speech. The worldwide platform of human voice creates possibilities to assail other users without facing any consequences, and flout social etiquettes, resulting in an inevitable increase of hate speech. Nowadays, English hate speech detection is a popular research area, but the prevalence of implicit hate content in regional languages desire effective language-independent models. The proposed research is the first unsupervised Hindi and Bengali hate content detection framework consisting of three significant concepts: HateCircle, hate tweet classification, and code-switch data preparation algorithms. The novel HateCircle method is proposed to detect hate orientation for each term by co-occurrence patterns of words, contextual semantics, and emotion analysis. The efficient multiclass hate tweet classification algorithm is proposed with parts of speech tagging, Euclidean distance, and the Geometric median methods. The detection of hate content is more efficient in the native script compared to the Roman script, so the transliteration algorithm is also proposed for code-switch data preparation. The experimentation evaluates the combination of various lexicons with our enriched hate lexicon that achieves a maximum of 0.74 F1-score for the Hindi and 0.88 F1-score for the Bengali datasets. The novel HateCircle and hate tweet detection framework evaluates with our proposed parts of speech tagging and Geometric median detection methods. Results reveal that HateCircle and hate tweet detection framework also achieves a maximum of 0.73 accuracy for the Hindi and 0.78 accuracy for the Bengali dataset. The experiment results signify that contextual semantic hate speech detection research with a language-independency feature offsets the growth of implicit abusive text in social media.

Detecting hate speech in memes: a review

Detecting hate speech in memes: a review

Hate Speech Classifiers Learn Normative Social Stereotypes

AbstractSocial stereotypes negatively impact individuals’ judgments about different groups and may have a critical role in understanding language directed toward marginalized groups. Here, we assess the role of social stereotypes in the automated detection of hate speech in the English language by examining the impact of social stereotypes on annotation behaviors, annotated datasets, and hate speech classifiers. Specifically, we first investigate the impact of novice annotators’ stereotypes on their hate-speech-annotation behavior. Then, we examine the effect of normative stereotypes in language on the aggregated annotators’ judgments in a large annotated corpus. Finally, we demonstrate how normative stereotypes embedded in language resources are associated with systematic prediction errors in a hate-speech classifier. The results demonstrate that hate-speech classifiers reflect social stereotypes against marginalized groups, which can perpetuate social inequalities when propagated at scale. This framework, combining social-psychological and computational-linguistic methods, provides insights into sources of bias in hate-speech moderation, informing ongoing debates regarding machine learning fairness.

ArHateDetector: detection of hate speech from standard and dialectal Arabic Tweets

Hate speech has become a phenomenon on social media platforms, such as Twitter. These websites and apps that were initially designed to facilitate our expression of free speech, are sometimes being used to spread hate towards each other. In the Arab region, Twitter is a very popular social media platform and thus the number of tweets that contain hate speech is increasing rapidly. Many tweets are written either in standard, dialectal Arabic, or mix. Existing work on Arabic hate speech are targeted towards either standard or single dialectal text, but not both. To fight hate speech more efficiently, in this paper, we conducted extensive experiments to investigate Arabic hate speech in tweets. Therefore, we propose a framework, called arHateDetector, that detects hate speech in the Arabic text of tweets. The proposed arHateDetector supports both standard and several dialectal Arabic. A large Arabic hate speech dataset, called arHateDataset, was compiled from several Arabic standard and dialectal tweets. The tweets are preprocessed to remove the unwanted content. We investigated the use of recent machine learning and deep learning models such as AraBERT to detect hate speech. All classification models used in the investigation are trained with the compiled dataset. Our experiments shows that AraBERT outperformed the other models producing the best performance across seven different datasets including the compiled arHateDataset with an accuracy of 93%. CNN and LinearSVC produced 88% and 89% respectively.

Hope speech detection in Spanish

In recent years, systems have been developed to monitor online content and remove abusive, offensive or hateful content. Comments in online social media have been analyzed to find and stop the spread of negativity using methods such as hate speech detection, identification of offensive language or detection of abusive language. We define hope speech as the type of speech that is able to relax a hostile environment and that helps, gives suggestions and inspires for good to a number of people when they are in times of illness, stress, loneliness or depression. Detecting it automatically, in order to give greater diffusion to positive comments, can have a very significant effect when it comes to fighting against sexual or racial discrimination or when we intend to foster less bellicose environments. In this article we perform a complete study on hope speech, analyzing existing solutions and available resources. In addition, we have generated a quality resource, SpanishHopeEDI, a new Spanish Twitter dataset on LGBT community, and we have conducted some experiments that can serve as a baseline for further research.

Deep feature fusion for hate speech detection: a transfer learning approach

Deep feature fusion for hate speech detection: a transfer learning approach