Hebrew offensive language taxonomy and dataset

The growing use of social media has increased online harassment, cyberhate, and the use of offensive language. This poses significant challenges for effectively detecting and addressing such issues. Natural Language Processing (NLP) has seen considerable advancements; however, automatically identifying offensive language remains a complex task due to the ambiguous and informal nature of user-generated content and the social context in which it occurs. In this thesis, our goal is to develop methods for automatic detection of offensive language in social media. Multiple classification algorithms, including Multinomial Naive Bayes, Gaussian Naive Bayes, SVM, Logistic Regression, and LSTM, are implemented and evaluated. Key measures including accuracy, F1 score, and AUC score are used to evaluate how well these algorithms work. Results show that the Random Forest Classifier obtains an AUC score of 0.65 and an accuracy of 0.82 without word2vec. On the other hand, LSTM demonstrates a competitive AUC score of 0.78 when compared to the Random Forest Classifier. These findings provide insights into the effectiveness of different algorithms for offensive language detection. The research contributes to the field by providing valuable tools and insights to enhance Turkish language processing and prioritize online safety, particularly in combating cyberbullying and fostering a tolerant online environment. The findings also pave the way for future research endeavors in natural language processing and have practical implications for protecting individuals and promoting a secure online space.

Social media has changed the world and play an important role in people lives. Social media platforms like Twitter, Facebook and YouTube create a new dimension of communication by providing channels to express and exchange ideas freely. Although the evolution brings numerous benefits, the dynamic environment and the allowable of anonymous posts could expose the uglier side of humanity. Irresponsible people would abuse the freedom of speech by aggressively express opinion or idea that incites hatred. This study performs hate speech and offensive language detection. The problem of this task is there is no clear boundary between hate speech and offensive language. In this study, a selected new features set is proposed for detecting hate speech and offensive language. Using Twitter dataset, the experiments are performed by considering the combination of word n-gram and enhanced syntactic n-gram. To reduce the feature set, filter-embedded combining feature selection is used. The experimental results indicate that the combination of word n-gram and enhanced syntactic n-gram with feature selection to classify the data into three classes: hate speech, offensive language or neither could give good performance. The result reaches 91% for accuracy and the averages of precision, recall and F1.

Software Engineering (SE) communities such as Stack Overflow have become unwelcoming, particularly through members' use of offensive language. Research has shown that offensive language drives users away from active engagement within these platforms. This work aims to explore this issue more broadly by investigating the nature of offensive language in comments posted by users in four prominent SE platforms - GitHub, Gitter, Slack and Stack Overflow (SO). It proposes an approach to detect and classify offensive language in SE communities by adopting natural language processing and deep learning techniques. Further, a Conflict Reduction System (CRS), which identifies offence and then suggests what changes could be made to minimize offence has been proposed. Beyond showing the prevalence of offensive language in over 1 million comments from four different communities which ranges from 0.07% to 0.43%, our results show promise in successful detection and classification of such language. The CRS system has the potential to drastically reduce manual moderation efforts to detect and reduce offence in SE communities.

The pervasiveness of offensive language on social media emphasizes the necessity of automated systems for identifying and categorizing content. To ensure a more secure online environment and improve communication, effective identification and categorization of this content is essential. However, existing research encounters challenges such as limited datasets and biased model performance, hindering progress in this domain. To address these challenges, this research presents a comprehensive framework that simplifies the utilization of support vector machines (SVM), random forest (RF) and artificial neural networks (ANN). The proposed methodology yields notable gains in offensive language detection, automatic categorization of offensiveness, and offense target identification tasks by utilizing the Offensive Language Identification Dataset (OLID). The simulation results indicate that SVM performs exceptionally well, exhibiting excellent accuracy scores (77%, 88%, and 68%), precision scores (76%, 87%, and 67%), F1 scores (57%, 88%, and 68%), and recall rates (45%, 88%, and 68%), proving to be practically successful in identifying and moderating offensive content on social media. By applying sophisticated preprocessing and meticulous hyperparameter tuning, our model outperforms some earlier research in detecting and categorizing offensive language tasks.

Offensive language is a toxic content that invades the web. It is characterized by being more than a negative reaction towards a certain event or person, since it is sharp and rude. Hate speech may be implied within an offensive language if it targets a specific part of a society, a religion, a political part, a race, a background or a specific gender. The spread of such languages over the web may cause psychological effects, hate, and war and can destruct a whole community. Preventing such a virus from spreading may help to prevent many social and political disastrous effects. In this paper, we aim to detect offensive language in social media reviews. Although detecting offensive language is a task similar to sentiment analysis, the targets of each task differ, however a transfer learning between them is possible. The offensive language detection system will be based on the same characteristics as a sentiment analysis system. We enhance the classification by performing ensemble learning that addresses the outputs of many classifiers. In order to detect hate speech, we based experiments on named entities. All along experiments, we perform both supervised and unsupervised approaches, and we also vary the lexicons based on which we generate features. The herein used lexicons are either internal extracted manually or automatically from the training corpus by the authors of this paper or external collected and filtered by our system.

Social media texts such as blog posts, comments, and tweets often contain offensive languages including racial hate speech comments, personal attacks, and sexual harassment. Detecting inappropriate use of language is, therefore, of utmost importance for the safety of the users as well as for suppressing hateful conduct and aggression. Existing approaches to this problem are mostly available for resource-rich languages such as English and German. In this paper, we characterize the offensive language in Nepali, a low-resource language, highlighting the challenges that need to be addressed for processing Nepali social media text. We also present experiments for detecting offensive language using supervised machine learning. Besides contributing the first baseline approaches of detecting offensive language in Nepali, we also release human annotated data sets to encourage future research on this crucial topic.

Machine translation (MT) is widely used to translate content on social media platforms aiming to improve accessibility. A great part of the content circulated on social media is user-generated and often contains non-standard spelling, hashtags, and emojis that pose challenges to MT systems. This leads to many mistranslated instances that are presented to users of these platforms, hindering their understanding of content written in other languages. In this paper, we investigate the impact of MT on offensive language identification. We pose that MT and potential mistranslations have an important and mostly under-explored impact on social media tasks such as sentiment analysis and offensive language identification. We create MT-Offense, a novel dataset containing English originals and translations in Arabic, Hindi, Marathi, Sinhala, and Spanish produced by multiple open-access Neural Machine Translation systems. We evaluate the performance of various offensive language models on both original and MT content in different training and test set combinations. We report the F1 scores of the models. Our results show that (1) offensive language identification models perform better on original data than on MT data, and (2) the use of MT data in training helps models better identify offensive language in MT content compared to models trained exclusively on original data.

Hate speech detection is an important issue in sentiment analysis and natural language processing. This study aims to improve the effectiveness of hate speech detection in English text using the BERT model, along with modified preprocessing techniques to enhance the F1-score. The dataset, sourced from Kaggle, contains English text with hate speech content. Evaluation results show a significant improvement in the model's accuracy and overall text classification performance. The BERT model achieved 89.11% accuracy on test data, correctly predicting 85 out of 95 samples. While the model excels at classifying offensive text with around 95% accuracy, it struggles to distinguish between hate and offensive text, with some confusion between neither and offensive categories. The classification report shows F1-scores of 0.43 for the hate class, 0.94 for the offensive class, and 0.84 for the neither class, with a weighted average F1-score of 0.89 and a macro average of 0.73. These results indicate that the BERT model delivers solid performance in detecting hate speech, though there is room for improvement, particularly in distinguishing certain classes.

Large language models (LLMs) demonstrate impressive capabilities across a range of natural language processing (NLP) tasks. However, they are highly sensitive to prompt design, which significantly affects their ability to align outputs with user intent. Poorly crafted prompts can result in misleading or irrelevant responses. Nevertheless, selecting the most effective prompt from several candidates remains an open challenge. Despite the growing importance of prompt engineering, there is no comprehensive framework to systematically evaluate prompts across multiple dimensions, such as similarity, performance, efficiency, and consistency, particularly in scenarios where performance can be traded off against computational cost or consistency. In this study, we propose a novel scoring framework to evaluate handcrafted prompts across four essential dimensions: Similarity, performance, efficiency (measured by latency, input tokens, and output tokens), and consistency. Considering Arabic, a relatively low-resource, morphologically rich language, as a case study, we evaluated this framework on six diverse text classification tasks: Dialect identification, sentiment analysis, offensive language detection, stance detection, emotion detection, and sarcasm detection. Our methodology assesses prompts across multiple LLMs (GPT-4o mini, LLaMA, ALLAM, and Claude 3.5 Haiku), providing valuable insights into model-specific and task-specific performance patterns. Results demonstrate that no single prompt universally excels across all dimensions; rather, optimal prompts vary based on specific task requirements and evaluation priorities. The proposed framework enables the identification of the most effective prompts for each application context while revealing important trade-offs between performance metrics. By addressing the unique challenges of Arabic NLP, this research not only advances prompt engineering for underrepresented languages but also provides a systematic and adaptable methodology for prompt evaluation that can enhance LLM performance across a range of linguistic contexts, diverse domains, tasks, and various model architectures.

Offensive language identification is to detect the hurtful tweets, derogatory comments, swear words on social media. As an emerging growth of social media communication, offensive language detection has received more attention in the last years; we focus to perform the task on English, Danish and Greek. We have investigated which can be effect more on pre-trained models BERT (Bidirectional Encoder Representation from Transformer) and Machine Learning Approaches. Our investigation shows the difference performance between the three languages and to identify the best performance is evaluated by the classification algorithms. In the shared task SemEval-2020, our team SSN_NLP_MLRG submitted for three languages that are Subtasks A, B, C in English, Subtask A in Danish and Subtask A in Greek. Our team SSN_NLP_MLRG obtained the F1 Scores as 0.90, 0.61, 0.52 for the Subtasks A, B, C in English, 0.56 for the Subtask A in Danish and 0.67 for the Subtask A in Greek respectively.

Social networks facilitate communication between people from all over the world. Unfortunately, the excessive use of social networks leads to the rise of antisocial behaviors such as the spread of online offensive language, cyberbullying (CB), and hate speech (HS). Therefore, abusive offensive and hate detection become a crucial part of cyberharassment. Manual detection of cyberharassment is cumbersome, slow, and not even feasible in rapidly growing data. In this study, we addressed the challenges of automatic detection of the offensive tweets in the Arabic language. The main contribution of this study is to design and implement an intelligent prediction system encompassing a two-stage optimization approach to identify and classify the offensive from the non-offensive text. In the first stage, the proposed approach fine-tuned the pre-trained word embedding models by training them for several epochs on the training dataset. The embeddings of the vocabularies in the new dataset are trained and added to the old embeddings. While in the second stage, it employed a hybrid approach of two classifiers, namely XGBoost and SVM, and a genetic algorithm (GA) to mitigate the drawback of the classifiers in finding the optimal hyperparameter values to run the proposed approach. We tested the proposed approach on Arabic Cyberbullying Corpus (ArCybC), which contains tweets collected from four Twitter domains: gaming, sports, news, and celebrities. The ArCybC dataset has four categories: sexual, racial, intelligence, and appearance. The proposed approach produced superior results, in which the SVM algorithm with the Aravec SkipGram word embedding model achieved an accuracy rate of 88.2% and an F1-score rate of 87.8%.

Since cyberbullying impacts both individual victims and entire society, research on abusive language and its detection has attracted attention in recent years. Because social media sites like Facebook, Instagram, Twitter, and others are so widely accessible, hate speech, bullying, sexism, racism, aggressive material, harassment, poisonous comments, and other types of abuse have all substantially increased. Due to the critical requirement to detect, regulate, and limit the spread of harmful content on social networking sites, we conducted this study to automate the detection of offensive language or cyberbullying. We created a new Arabic balanced data set to be used in the offensive detection process because having a balanced data set for a model would result in improved accuracy models. Recently, the performance of single classifiers has been improved using ensemble machine learning. The purpose of this study is to examine the effectiveness of several single and ensemble machine learning algorithms in identifying Arabic text that contains foul language and cyberbullying. Applying them to three Arabic datasets, we have selected three machine learning classifiers and three ensemble models for this aim. Two of them are offensive datasets that are readily accessible in the public, while the third one was created. The results showed that the single learner machine learning strategy is inferior to the ensemble machine learning methodology. Voting performs is the best performing trained ensemble machine learning classifier, outperforming the best single learner classifier (65.1%, 76.2%, and 98%) for the same datasets with accuracy scores of (71.1%, 76.7%, and 98.5%) for each of the three datasets used. Finally, we improve the voting technique’s performance through hyperparameter tuning on the Arabic cyberbullying data set.

414 LANGUAGE, VOLUME 67, NUMBER 2 (1991) Brixhe & Lejeune's 1984 Corpus des inscriptions paléo-phrygiennes. Since Phrygian is both closely related to Greek and preserved in part through Greek testimony, N has chosen to present his account as a history of PhrygianGreek linguistic contact. This procedure may explain his vagueness in discussing shared features of the two languages: although he calls them 'common innovations' (10), he seems elsewhere to regard them as the result of borrowing during a hypothesized period of close contact among Greek, Phrygian, and several other IE languages in the south Balkans (5-6). Some of these shared features are quite striking—e.g. factitives in *-oyô like Gk. orthöö 'straighten'. Phr. kakoioi (11, 13), and laryngeal vocalization followed by umlaut in Gk. ónoma, Phr. onoman = [b] not [bh], this development is trivial and proves only that Phrygian did not share the Greek change to [ph] etc.). Pp. 6-10: Common retentions of Greek and Phrygian (e.g. the augment and the thirdperson present middle ending -toi). Pp. 10-13: Common innovations of Greek and Phrygian (e.g. agent nouns in -tä, Gk. autos 'self = Phr. autos 'particular'). Pp. 13-16: Inherited lexical material of Phrygian which is lost or only marginally preserved in Greek. Pp. 16-18: Sometime after the hypothesized period of south Balkan close contact with Greek, the Phrygians migrated to Anatolia, where they are attested in historical times. Contact with speakers ofAnatolian languages led to borrowings into Phrygian (mainly names, e.g. Midas). Pp. 18-19: Greek expansion in the first millennium led to contact and hence Greek loans into Phrygian. Pp. 19-22: Direct attestation of Phrygian lost from the third century B.C.E. through the first century CE. Pp. 22-23: Neo-Phrygian attested in grave inscriptions, still in part uninterpretable . Although Phrygian is unlikely ever to play a significant role in IE historical linguistics, it does, as N remarks in conclusion (23), repay time spent applying to it our knowledge of IE. N's essay is a valuable contribution to the project of understanding Phrygian. [Andrew Garrett, University of Texas, Austin.] Natural language processing technologies in Artificial Intelligence: The science and industry perspective. By Klaus K. Obermeier. Chichester : Ellis Horwood, 1989. Pp. 263. Natural language processing (NLP) is the area of research concerned with the many aspects of 'natural' language communication with computers . In this book Obermeier provides an introduction to NLP from both commercial and academic perspectives, primarily aimed at businessmen and data-processing professionals. His presentation is nontechnical, though very comprehensive . Linguists will find the book an accessible introduction to the practical issues in NLP, which may help to clarify the direction and emphasis in this area of language research. The book's organization is excellent, making it very easy to read, although the typographic and editorial quality are not the best. The twelve chapters are divided into three sections: 'NLP in perspective' (17-92). which provides an introduction to the structure of NLP systems and the approaches and formalisms used to implement them; 'Applications' (93-185), in which six major applications of NLP are described: and 'Bottom line' (187-232), where O presents his own evaluation of the state of NLP research and current systems. A good list of references (239-59) is also included. Much of the material in the first section will be familiar to linguists, since the design of NLP systems is largely based on formal linguistic analyses in the areas ofphonology, morphology, syntax, and semantics. In addition, methods for knowledge representation and reasoning drawn from Artificial Intelligence are explained in relation to their relevance to NLP development. This nontechnical overview of the techniques used in NLP is up-to-date. It includes a discussion of parsing methods ranging from contextfree parsing, unification-based parsing, and ATNs to such knowledge-representation schemes from AI as production rules, semantic networks, and frames, to the new connectionist models. BOOK NOTICES 415 The second section comprises the largest part of the book, with one chapter devoted to each of the applications addressed: natural-language interfaces, machine translation, text processing/ understanding, text generation, speech processing , and writing aids. For each application there is a brief...

Objectives: To provide an organised literature on the detection of Abusive language on Twitter using natural language processing (NLP). Methods: In this study, the survey has been conducted on different methods and research conducted on the types of Abusive language used in social media, why it is important? How it has been detected in real time social media platforms and the performance metrics that are used by researchers in evaluating the performance of the detection of abusive language on Twitter by the users. Results: Giving an organised review of past methodologies, including methods, important features and core algorithms, this study arranges and depicts the present condition about this area. The study also talks about the intricacy of hate speech idea which is characterised in numerous stages ad settings. This area of study has an obvious potential for societal effect, especially in digital media and online networks. A crucial step in propelling automatic hate speech detection is the advancement and systemisation of common assets, for example, clarified data sets in numerous dialects, rules, and calculations. Conclusion: This survey study contains all the relevant references related to detection of abusive language on social media using NLP and machine learning methods. Ultimately, it can be as source of references to the other researchers in finding the literatures that are relevant to their research area in the detection of Abusive language on Twitter. Keywords: Abusive Language, Natural Language Processing, Social media analysis, Text Classification and Analysis

The widespread of offensive content online, such as hate speech and cyber-bullying, is a global phenomenon. This has sparked interest in the artificial intelligence (AI) and natural language processing (NLP) communities, motivating the development of various systems trained to detect potentially harmful content automatically. These systems require annotated datasets to train the machine learning (ML) models. However, with a few notable exceptions, most datasets on this topic have dealt with English and a few other high-resource languages. As a result, the research in offensive language identification has been limited to these languages. This paper addresses this gap by tackling offensive language identification in Sinhala, a low-resource Indo-Aryan language spoken by over 17 million people in Sri Lanka. We introduce the Sinhala Offensive Language Dataset (SOLD) and present multiple experiments on this dataset. SOLD is a manually annotated dataset containing 10,000 posts from Twitter annotated as offensive and not offensive at both sentence-level and token-level, improving the explainability of the ML models. SOLD is the first large publicly available offensive language dataset compiled for Sinhala. We also introduce SemiSOLD, a larger dataset containing more than 145,000 Sinhala tweets, annotated following a semi-supervised approach.