Search

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. [ABSTRACT FROM AUTHOR]

The growing ubiquity of online anonymity has significantly transformed the dynamics of participation and collaboration across digital platforms, especially through social media. While this veil of anonymity enables the free expression of personal opinions, it also leads to the spread of negative content, including offensive language. The detection and mitigation of such offensive language pose a considerable challenge and represent a dynamic research area. This paper introduces and evaluates a new ensemble learning approach designed for the Arabic offensive language detection task. The proposed approach combines three distinct models: the pretrained Bidirectional Encoder Representations from Transformers (BERT) model, BERT combined with Global Average and Global Max layers, and BERT augmented with pooled stacked Bidirectional Long Short-Term Memory (Bi-LSTMs). Outperforming the performance of the baseline OffensEval2020 winner in "SemEval-2020 Task 12 on Multilingual Offensive Language Identification in Social Media", our model achieved a 90.97% F1-score on the original Arabic OffensEval2020 dataset and an enhanced 94.56% F1-score on the augmented dataset. [ABSTRACT FROM AUTHOR]

Social media serves as a platform for netizens to stay informed and express their opinions through the Internet. Currently, the social media discourse environment faces a significant security threat—offensive comments. A group of users posts comments that are provocative, discriminatory, and objectionable, intending to disrupt online discussions, provoke others, and incite intergroup conflict. These comments undermine citizens' legitimate rights, disrupt social order, and may even lead to real-world violent incidents. However, current automatic detection of offensive language primarily focuses on a few high-resource languages, leaving low-resource languages, such as Malay, with insufficient annotated corpora for effective detection. To address this, we propose a zero-shot, cross-language unsupervised offensive language detection (OLD) method using a dual-branch mBERT transfer approach. Firstly, using the multi-language BERT (mBERT) model as the foundational language model, the first network branch automatically extracts features from both source and target domain data. Subsequently, Sinkhorn distance is employed to measure the discrepancy between the source and target language feature representations. By estimating the Sinkhorn distance between the labeled source language (e.g., English) and the unlabeled target language (e.g., Malay) feature representations, the method minimizes the Sinkhorn distance adversarially to provide more stable gradients, thereby extracting effective domain-shared features. Finally, offensive pivot words from the source and target language training sets are identified. These pivot words are then removed from the training data in a second network branch, which employs the same architecture. This process constructs an auxiliary OLD task. By concealing offensive pivot words in the training data, the model reduces overfitting and enhances robustness to the target language. In the end-to-end framework training, the combination of cross-lingual shared features and independent features culminates in unsupervised detection of offensive speech in the target language. The experimental results demonstrate that employing cross-language model transfer learning can achieve unsupervised detection of offensive content in low-resource languages. The number of labeled samples in the source language is positively correlated with transfer performance, and a greater similarity between the source and target languages leads to better transfer effects. The proposed method achieves the best performance in OLD on the Malay dataset, achieving an F1 score of 80.7%. It accurately identifies features of offensive speech, such as sarcasm, mockery, and implicit expressions, and showcases strong generalization and excellent stability across different target languages. [ABSTRACT FROM AUTHOR]

With the proliferation of social networks and their impact on human life, one of the rising problems in this environment is the rise in verbal and written insults and hatred. As one of the significant platforms for distributing text-based content, Twitter frequently publishes its users' abusive remarks. Creating a model that requires a complete collection of offensive sentences is the initial stage in recognizing objectionable phrases. In addition, despite the abundance of resources in English and other languages, there are limited resources and studies on identifying hateful and offensive statements in Persian. In this study, we compiled a 38K-tweet dataset of Persian Hate and Offensive language using keyword-based data selection strategies. A Persian offensive lexicon and nine hatred target group lexicons were gathered through crowdsourcing for this purpose. The dataset was annotated manually so that at least two annotators investigated tweets. In addition, for the purpose of analyzing the effect of used lexicons on language model functionality, we employed two assessment criteria (FPED and pAUCED) to measure the dataset's potential bias. Then, by configuring the dataset based on the results of the bias measurement, we mitigated the effect of words' bias in tweets on language model performance. The results indicate that bias is significantly diminished, while less than a hundredth reduced the F1 score. [ABSTRACT FROM AUTHOR]

With the advent of Web 2.0, various platforms and tools have been developed to allow internet users to express their opinions and thoughts on diverse topics and occurrences. Nevertheless, certain users misuse these platforms by sharing hateful and offensive speeches, which has a negative impact on the mental health of internet society. Thus, the detection of offensive language has become an active area of research in the field of natural language processing. Rapidly detecting offensive language on the internet and preventing it from spreading is of great practical significance in reducing cyberbullying and self-harm behaviors. Despite the crucial importance of this task, limited work has been done in this field for non-English languages such as Arabic. Therefore, in this paper, we aim to improve the results of Arabic offensive language detection without the need for laborious preprocessing or feature engineering work. To achieve this, we combine the bidirectional encoder representations from transformers (BERT) model model with a bidirectional gated recurrent unit (BiGRU) layer to further enhance the extracted context and semantic features. The experiments were conducted on the Arabic dataset provided by the SemEval 2020 Task 12. The evaluation results show the effectiveness of our model compared to the baseline and related work models by achieving a macro F1- score of 93.16%. [ABSTRACT FROM AUTHOR]

Copyright of Procesamiento del Lenguaje Natural is the property of Sociedad Espanola para el Procesamiento del Lenguaje Natural and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

The advent of online platforms and services has revolutionized communication, enabling users to share opinions and ideas seamlessly. However, this convenience has also brought about a surge in offensive and harmful language across various communication mediums. In response, social platforms have turned to automated methods to identify offensive content. A critical research question emerges when investigating the role of specific text spans within comments in conveying offensive characteristics. This paper conducted a comprehensive investigation into detecting offensive text spans in Romanian language comments using Transformer encoders and Large Language Models (LLMs). We introduced an extensive dataset of 4800 Romanian comments annotated with offensive text spans. Moreover, we explored the impact of varying model sizes, architectures, and training data volumes on the performance of offensive text span detection, providing valuable insights for determining the optimal configuration. The results argue for the effectiveness of BERT pre-trained models for this span-detection task, showcasing their superior performance. We further investigated the impact of different sample-retrieval strategies for few-shot learning using LLMs based on vector text representations. The analysis highlights important insights and trade-offs in leveraging LLMs for offensive-language-detection tasks. [ABSTRACT FROM AUTHOR]

Social media serves as a platform for netizens to stay informed and express their opinions through the Internet. Currently, the social media discourse environment faces a significant security threat—offensive comments. A group of users posts comments that are provocative, discriminatory, and objectionable, intending to disrupt online discussions, provoke others, and incite intergroup conflict. These comments undermine citizens’ legitimate rights, disrupt social order, and may even lead to real-world violent incidents. However, current automatic detection of offensive language primarily focuses on a few high-resource languages, leaving low-resource languages, such as Malay, with insufficient annotated corpora for effective detection. To address this, we propose a zero-shot, cross-language unsupervised offensive language detection (OLD) method using a dual-branch mBERT transfer approach. Firstly, using the multi-language BERT (mBERT) model as the foundational language model, the first network branch automatically extracts features from both source and target domain data. Subsequently, Sinkhorn distance is employed to measure the discrepancy between the source and target language feature representations. By estimating the Sinkhorn distance between the labeled source language (e.g., English) and the unlabeled target language (e.g., Malay) feature representations, the method minimizes the Sinkhorn distance adversarially to provide more stable gradients, thereby extracting effective domain-shared features. Finally, offensive pivot words from the source and target language training sets are identified. These pivot words are then removed from the training data in a second network branch, which employs the same architecture. This process constructs an auxiliary OLD task. By concealing offensive pivot words in the training data, the model reduces overfitting and enhances robustness to the target language. In the end-to-end framework training, the combination of cross-lingual shared features and independent features culminates in unsupervised detection of offensive speech in the target language. The experimental results demonstrate that employing cross-language model transfer learning can achieve unsupervised detection of offensive content in low-resource languages. The number of labeled samples in the source language is positively correlated with transfer performance, and a greater similarity between the source and target languages leads to better transfer effects. The proposed method achieves the best performance in OLD on the Malay dataset, achieving an F1 score of 80.7%. It accurately identifies features of offensive speech, such as sarcasm, mockery, and implicit expressions, and showcases strong generalization and excellent stability across different target languages.

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 in social media affects the social experience of individuals and groups and hurts social harmony and moral values. Therefore, in recent years, the problem of offensive language detection has attracted the attention of many researchers. However, the primary research currently focuses on detecting English offensive language, while few studies on the Chinese language exist. In this paper, we propose an innovative approach to detect Chinese offensive language. First, unlike previous approaches, we utilized both RoBERTa's sentence-level and word-level embedding, combining the sentence embedding and word embedding of RoBERTa's model, bidirectional GRU, and multi-head self-attention mechanism. This feature fusion allows the model to consider sentence-level and word-level semantic information at the same time so as to capture the semantic information of Chinese text more comprehensively. Second, by concatenating the output results of multi-head attention with RoBERTa's sentence embedding, we achieved an efficient fusion of local and global information and improved the representation ability of the model. The experiments showed that the proposed model achieved 82.931% accuracy and 82.842% F1-score in Chinese offensive language detection tasks, delivering high performance and broad application potential. [ABSTRACT FROM AUTHOR]

With the increasing use of social media with its ability for users to share comments immediately, the extent of a system to identify offensive content has become a necessity in all languages. Due to the lack of publicly available resources on offensive language identification for Farsi, which has more than 110 million speakers, we present Pars-OFF, a three-layered annotated corpus for offensive language detection in Farsi to fill the existing gap. The introduced corpus contains 10,563 data samples. The tweets have been collected with a combination of similarity-based and keyword-based data selection techniques to avoid severe unbalancedness. Additionally, as a baseline, this article reports the performance of the traditional machine learning approaches and Transformer based models over the Pars-OFF dataset. The best performance was obtained by the BERT+fastText model, yielding the F1-Macro score of 89.57. [ABSTRACT FROM AUTHOR]

Social media platforms have become inundated with offensive language. This issue must be addressed for the growth of online social networks (OSNs) and a healthy online environment. While significant research has been devoted to identifying toxic content in major languages like English, this remains an open area of research in the low-resource Pashto language. This study aims to develop an AI model for the automatic detection of offensive textual content in Pashto. To achieve this goal, we have developed a benchmark dataset called the Pashto Offensive Language Dataset (POLD), which comprises tweets collected from Twitter and manually classified into two categories: "offensive" and "not offensive". To discriminate these two categories, we investigated the classic deep learning classifiers based on neural networks, including CNNs and RNNs, using static word embeddings: Word2Vec, fastText, and GloVe as features. Furthermore, we examined two transfer learning approaches. In the first approach, we fine-tuned the pre-trained multilingual language model, XLM-R, using the POLD dataset, whereas, in the second approach, we trained a monolingual BERT model for Pashto from scratch using a custom-developed text corpus. Pashto BERT was then fine-tuned similarly to XLM-R. The performance of all the deep learning and transformer learning models was evaluated using the POLD dataset. The experimental results demonstrate that our pre-trained Pashto BERT model outperforms the other models, achieving an F1-score of 94.34% and an accuracy of 94.77%. [ABSTRACT FROM AUTHOR]

Controversy measurement on social media plays an important part in understanding public opinion. Various topics are frequently hotly debated on social media platforms including Twitter and Sina Weibo. People sometimes use offensive or sarcastic language to convey their opinions about a topic or a source post, which might spark heated discussions and controversy towards related topics. Recent researches take controversy detection as a binary classification problem with two labels: controversy or non-controversy. The reason might be lacking a comprehensive understanding of why the controversy happened and a specific imagination of how it will be used in the downstream tasks. However, we believe that the degree of controversy courted by posts or topics in a real scenario varied. And fine-grained measurement of controversy will be beneficial to public sentiment identification, influence assessment and other social network analysis tasks. We also notice that the existing benchmarks of controversy detection are not applicable for fine-grained topic-level controversy measurement. In this paper, we present ProsCons, a large-scale comprehensive Chinese dataset that includes 245 topics and 32,667 posts with pro, con or neutral stances. Based on that, we design a controversy measurement framework for measuring the controversy intensity that topics sparked. This framework considers the degree of antagonism in terms of stance and sentiment, as well as the irrational degree (offensive or sarcasm) of a post to compute a controversy intensity. ProsCons provides a new benchmark for Chinese stance detection, offensive language and sarcasm detection, contributing to the multi-task learning of them. We conduct extensive experiments on ProsCons and provide baselines for these tasks. The experimental results highlight the challenges of the aforementioned tasks based on the ProsCons. [ABSTRACT FROM AUTHOR]

Offensive communications have invaded social media content. One of the most effective solutions to cope with this problem is using computational techniques to discriminate offensive content. Moreover, social media users are from linguistically different communities. This study aims to tackle the Multilingual Offensive Language Detection (MOLD) task using transfer learning models and the fine-tuning phase. We propose an effective approach based on the Bidirectional Encoder Representations from Transformers (BERT) that has shown great potential in capturing the semantics and contextual information within texts. The proposed system consists of several stages: (1) Preprocessing, (2) Text representation using BERT models, and (3) Classification into two categories: Offensive and non-offensive. To handle multilingualism, we explore different techniques such as the joint-multilingual and translation-based ones. The first consists in developing one classification system for different languages, and the second involves the translation phase to transform all texts into one universal language then classify them. We conduct several experiments on a bilingual dataset extracted from the Semi-supervised Offensive Language Identification Dataset (SOLID). The experimental findings show that the translation-based method in conjunction with Arabic BERT (AraBERT) achieves over 93% and 91% in terms of F1-score and accuracy, respectively.

The proliferation of harmful content on social media affects a large part of the user community. Therefore, several approaches have emerged to control this phenomenon automatically. However, this is still a quite challenging task. In this paper, we explore the offensive language as a particular case of harmful content and focus our study in the analysis of keywords in available datasets composed of offensive tweets. Thus, we aim to identify relevant words in those datasets and analyze how they can affect model learning. For keyword extraction, we propose an unsupervised hybrid approach which combines the multi-head self-attention of BERT and a reasoning on a word graph. The attention mechanism allows to capture relationships among words in a context, while a language model is learned. Then, the relationships are used to generate a graph from what we identify the most relevant words by using the eigenvector centrality. Experiments were performed by means of two mechanisms. On the one hand, we used an information retrieval system to evaluate the impact of the keywords in recovering offensive tweets from a dataset. On the other hand, we evaluated a keyword-based model for offensive language detection. Results highlight some points to consider when training models with available datasets. [ABSTRACT FROM AUTHOR]

The advent of online platforms and services has revolutionized communication, enabling users to share opinions and ideas seamlessly. However, this convenience has also brought about a surge in offensive and harmful language across various communication mediums. In response, social platforms have turned to automated methods to identify offensive content. A critical research question emerges when investigating the role of specific text spans within comments in conveying offensive characteristics. This paper conducted a comprehensive investigation into detecting offensive text spans in Romanian language comments using Transformer encoders and Large Language Models (LLMs). We introduced an extensive dataset of 4800 Romanian comments annotated with offensive text spans. Moreover, we explored the impact of varying model sizes, architectures, and training data volumes on the performance of offensive text span detection, providing valuable insights for determining the optimal configuration. The results argue for the effectiveness of BERT pre-trained models for this span-detection task, showcasing their superior performance. We further investigated the impact of different sample-retrieval strategies for few-shot learning using LLMs based on vector text representations. The analysis highlights important insights and trade-offs in leveraging LLMs for offensive-language-detection tasks.

Offensive language in social media affects the social experience of individuals and groups and hurts social harmony and moral values. Therefore, in recent years, the problem of offensive language detection has attracted the attention of many researchers. However, the primary research currently focuses on detecting English offensive language, while few studies on the Chinese language exist. In this paper, we propose an innovative approach to detect Chinese offensive language. First, unlike previous approaches, we utilized both RoBERTa’s sentence-level and word-level embedding, combining the sentence embedding and word embedding of RoBERTa’s model, bidirectional GRU, and multi-head self-attention mechanism. This feature fusion allows the model to consider sentence-level and word-level semantic information at the same time so as to capture the semantic information of Chinese text more comprehensively. Second, by concatenating the output results of multi-head attention with RoBERTa’s sentence embedding, we achieved an efficient fusion of local and global information and improved the representation ability of the model. The experiments showed that the proposed model achieved 82.931% accuracy and 82.842% F1-score in Chinese offensive language detection tasks, delivering high performance and broad application potential.

In recent years, social media networks are emerging as a key player by providing platforms for opinions expression, communication, and content distribution. However, users often take advantage of perceived anonymity on social media platforms to share offensive or hateful content. Thus, offensive language has grown as a significant issue with the increase in online communication and the popularity of social media platforms. This problem has attracted significant attention for devising methods for detecting offensive content and preventing its spread on online social networks. Therefore, this paper aims to develop an effective Arabic offensive language detection model by employing deep learning and semantic and contextual features. This paper proposes a deep learning approach that utilizes the bidirectional long short-term memory (BiLSTM) model and domain-specific word embeddings extracted from an Arabic offensive dataset. The detection approach was evaluated on an Arabic dataset collected from Twitter. The results showed the highest performance accuracy of 0.93% with the BiLSTM model trained using a combination of domain-specific and agnostic-domain word embeddings.

Offensive communications have invaded social media content. One of the most effective solutions to cope with this problem is using computational techniques to discriminate offensive content. Moreover, social media users are from linguistically different communities. This study aims to tackle the Multilingual Offensive Language Detection (MOLD) task using transfer learning models and the fine-tuning phase. We propose an effective approach based on the Bidirectional Encoder Representations from Transformers (BERT) that has shown great potential in capturing the semantics and contextual information within texts. The proposed system consists of several stages: (1) Preprocessing, (2) Text representation using BERT models, and (3) Classification into two categories: Offensive and non-offensive. To handle multilingualism, we explore different techniques such as the joint-multilingual and translation-based ones. The first consists in developing one classification system for different languages, and the second involves the translation phase to transform all texts into one universal language then classify them. We conduct several experiments on a bilingual dataset extracted from the Semi-supervised Offensive Language Identification Dataset (SOLID). The experimental findings show that the translation-based method in conjunction with Arabic BERT (AraBERT) achieves over 93% and 91% in terms of F1-score and accuracy, respectively. [ABSTRACT FROM AUTHOR]

In recent years, social media networks are emerging as a key player by providing platforms for opinions expression, communication, and content distribution. However, users often take advantage of perceived anonymity on social media platforms to share offensive or hateful content. Thus, offensive language has grown as a significant issue with the increase in online communication and the popularity of social media platforms. This problem has attracted significant attention for devising methods for detecting offensive content and preventing its spread on online social networks. Therefore, this paper aims to develop an effective Arabic offensive language detection model by employing deep learning and semantic and contextual features. This paper proposes a deep learning approach that utilizes the bidirectional long short-term memory (BiLSTM) model and domain-specific word embeddings extracted from an Arabic offensive dataset. The detection approach was evaluated on an Arabic dataset collected from Twitter. The results showed the highest performance accuracy of 0.93% with the BiLSTM model trained using a combination of domain-specific and agnostic-domain word embeddings. [ABSTRACT FROM AUTHOR]

In recent years, unethical behavior in the cyber-environment has been revealed. The presence of offensive language on social media platforms and automatic detection of such language is becoming a major challenge in modern society. The complexity of natural language constructs makes this task even more challenging. Until now, most of the research has focused on resource-rich languages like English. Roman Urdu and Urdu are two scripts of writing the Urdu language on social media. The Roman script uses the English language characters while the Urdu script uses Urdu language characters. Urdu and Hindi languages are similar with the only difference in their writing script but the Roman scripts of both languages are similar. This study is about the detection of offensive language from the user's comments presented in a resource-poor language Urdu. We propose the first offensive dataset of Urdu containing user-generated comments from social media. We use individual and combined n-grams techniques to extract features at character-level and word-level. We apply seventeen classifiers from seven machine learning techniques to detect offensive language from both Urdu and Roman Urdu text comments. Experiments show that the regression-based models using character n-grams show superior performance to process the Urdu language. Character-level tri-gram outperforms the other word and character n-grams. LogitBoost and SimpleLogistic outperform the other models and achieve 99.2% and 95.9% values of F-measure on Roman Urdu and Urdu datasets respectively. Our designed dataset is publically available on GitHub for future research.

Harmful language is frequent in social media, in particular in spaces which are considered anonymous and/or allow free participation. In this paper, we analyze the language in a Telegram channel populated by followers of former US President Donald Trump. We seek to identify the ways in which harmful language is used to create a specific narrative in a group of mostly like-minded discussants. Our research has several aims. First, we create an extended taxonomy of potentially harmful language that includes not only hate speech and direct insults (which have been the focus of existing computational methods), but also other forms of harmful speech discussed in the literature. We manually apply this taxonomy to a large portion of the corpus, including the time period leading up to and the aftermath of the January 2021 US Capitol riot. Our data gives empirical evidence for harmful speech, such as in/out-group divisive language and the use of codes within certain communities, that have not often been investigated before. Second, we compare our manual annotations of harmful speech to several automatic methods for classifying hate speech and offensive language, namely list-based and machine-learning-based approaches. We find that the Telegram data sets still pose particular challenges for these automatic methods. Finally, we argue for the value of studying such naturally-occurring, coherent data sets for research on online harm and how to address it in linguistics and philosophy.

Copyright of Procesamiento del Lenguaje Natural is the property of Sociedad Espanola para el Procesamiento del Lenguaje Natural and its content may not be copied or emailed to multiple sites or posted to a listserv without the copyright holder's express written permission. However, users may print, download, or email articles for individual use. This abstract may be abridged. No warranty is given about the accuracy of the copy. Users should refer to the original published version of the material for the full abstract. (Copyright applies to all Abstracts.)

Platforms that feature user-generated content (social media, online forums, newspaper comment sections etc.) have to detect and filter offensive speech within large, fast-changing datasets. While many automatic methods have been proposed and achieve good accuracies, most of these focus on the English language, and are hard to apply directly to languages in which few labeled datasets exist. Recent work has therefore investigated the use of cross-lingual transfer learning to solve this problem, training a model in a well-resourced language and transferring to a less-resourced target language; but performance has so far been significantly less impressive. In this paper, we investigate the reasons for this performance drop, via a systematic comparison of pre-trained models and intermediate training regimes on five different languages. We show that using a better pre-trained language model results in a large gain in overall performance and in zero-shot transfer, and that intermediate training on other languages is effective when little target-language data is available. We then use multiple analyses of classifier confidence and language model vocabulary to shed light on exactly where these gains come from and gain insight into the sources of the most typical mistakes.

Harmful language is frequent in social media, in particular in spaces which are considered anonymous and/or allow free participation. In this paper, we analyze the language in a Telegram channel populated by followers of former US President Donald Trump. We seek to identify the ways in which harmful language is used to create a specific narrative in a group of mostly like-minded discussants. Our research has several aims. First, we create an extended taxonomy of potentially harmful language that includes not only hate speech and direct insults (which have been the focus of existing computational methods), but also other forms of harmful speech discussed in the literature. We manually apply this taxonomy to a large portion of the corpus, including the time period leading up to and the aftermath of the January 2021 US Capitol riot. Our data gives empirical evidence for harmful speech, such as in/outgroup divisive language and the use of codes within certain communities, that have not often been investigated before. Second, we compare our manual annotations of harmful speech to several automatic methods for classifying hate speech and offensive language, namely list-based and machine-learning-based approaches. We find that the Telegram data sets still pose particular challenges for these automatic methods. Finally, we argue for the value of studying such naturally-occurring, coherent data sets for research on online harm and how to address it in linguistics and philosophy. [ABSTRACT FROM AUTHOR]

The fight against offensive speech on the Internet necessitates increased efforts from linguistic analysis and artificial intelligence perspectives to develop countermeasures and preventive methods. Reliable predictions can only be obtained if these methods are exposed to a representative sample of the domain or environment under consideration. Datasets serve as the foundation for significant developments in this field because they are the main means of obtaining appropriate instances that reveal the multiple and varied faces of the offensive speech phenomenon. In this sense, we present Ar-PuFi, a dataset of offensive speech towards Public Figures in the Arabian community. With 24,071 comments collected from TV interviews with Egyptian celebrities belonging to six domains of public interest, Ar-PuFi is currently the largest Arabic dataset in terms of its category and size. The examples were annotated by three native speakers over the course of two months and are provided with two-class and six-class annotations based on the presence or absence of explicit and implicit offensive content. We evaluated the performance of a diverse set of classification models employing several text representations of actual examples (e.g., N-gram, TF/IDF, AraVec, and fastText), and AraBERT achieved the baseline for the new dataset in both offensive detection and group classification. Additionally, we apply the Pointwise Mutual Information (PMI) technique to comments within the target domain in order to derive a lexicon of offensive terms associated with each domain of ArPuFi. We further explored whether active learning (AL) or meta-learning (ML) frameworks could be used to reduce the labeling effort required for our dataset without affecting prediction quality and found that, though AL can reduce the amount of data annotations by 10% over the ML approach, neither approach requires less than about 70% of the full dataset to achieve baseline performance. Finally, we took advantage of the availability of relevant datasets and conducted a cross-domain experiment to back up our claims not only about the uniqueness of our dataset but also about the difficulty of adapting Arabic dialects against one another. [ABSTRACT FROM AUTHOR]

Social media platforms have become inundated with offensive language. This issue must be addressed for the growth of online social networks (OSNs) and a healthy online environment. While significant research has been devoted to identifying toxic content in major languages like English, this remains an open area of research in the low-resource Pashto language. This study aims to develop an AI model for the automatic detection of offensive textual content in Pashto. To achieve this goal, we have developed a benchmark dataset called the Pashto Offensive Language Dataset (POLD), which comprises tweets collected from Twitter and manually classified into two categories: "offensive" and "not offensive". To discriminate these two categories, we investigated the classic deep learning classifiers based on neural networks, including CNNs and RNNs, using static word embeddings: Word2Vec, fastText, and GloVe as features. Furthermore, we examined two transfer learning approaches. In the first approach, we fine-tuned the pre-trained multilingual language model, XLM-R, using the POLD dataset, whereas, in the second approach, we trained a monolingual BERT model for Pashto from scratch using a custom-developed text corpus. Pashto BERT was then fine-tuned similarly to XLM-R. The performance of all the deep learning and transformer learning models was evaluated using the POLD dataset. The experimental results demonstrate that our pre-trained Pashto BERT model outperforms the other models, achieving an F1-score of 94.34% and an accuracy of 94.77%., Competing Interests: The authors declare there are no competing interests., (©2023 Haq et al.)

Abstract We present the results of predictive modelling for the detection of anti-social behaviour in online communication in Arabic, such as comments which contain obscene or offensive words and phrases. We collected and labelled a large dataset of YouTube comments in Arabic which contains a broad range of both offensive and inoffensive comments. We used this dataset to train a Support Vector Machine classifier and experimented with combinations of word-level features, N-gram features and a variety of pre-processing techniques. We summarise the pre-processing steps and features that allow training a classifier which is more precise, with 90.05% accuracy, than classifiers reported by previous studies on Arabic text. [ABSTRACT FROM AUTHOR]

Platforms that feature user-generated content (social media, online forums, newspaper comment sections etc.) have to detect and filter offensive speech within large, fast-changing datasets. While many automatic methods have been proposed and achieve good accuracies, most of these focus on the English language, and are hard to apply directly to languages in which few labeled datasets exist. Recent work has therefore investigated the use of cross-lingual transfer learning to solve this problem, training a model in a well-resourced language and transferring to a less-resourced target language; but performance has so far been significantly less impressive. In this paper, we investigate the reasons for this performance drop, via a systematic comparison of pre-trained models and intermediate training regimes on five different languages. We show that using a better pre-trained language model results in a large gain in overall performance and in zero-shot transfer, and that intermediate training on other languages is effective when little target-language data is available. We then use multiple analyses of classifier confidence and language model vocabulary to shed light on exactly where these gains come from and gain insight into the sources of the most typical mistakes., Competing Interests: The authors declare that they have no competing interests., (© 2021 Pelicon et al.)