Bangladesh has seen a remarkable increase in its use of the Internet over the past two decades. There were more than 125 million Internet users in Bangladesh as of November 2022, according to the Bangladesh Telecommunication Regulatory Commission (BTRC) ^[1]. Additionally, with the help of the implementation of the Digital Bangladesh initiative ^[2], the vast majority of people in Chittagong ^[3], Bangladesh’s second-largest city, now have access to the Internet and actively use social media. According to a survey ^[4], the number of Facebook users in Bangladesh is the highest among social media (see Figure 1). Moreover, with the benefit of Unicode being widely used on most communication devices, such as tablets or smartphones, speakers of underrepresented languages, such as Chittagonian, can express their thoughts in their native languages and dialects. Many people in Chittagong now use social media on a regular basis, regularly using platforms like Facebook ^[5], imo ^[6], various blogs, and WhatsApp ^[7]. These platforms offer a venue where people can express themselves freely and informally. However, the pervasiveness of social media has also resulted in unfavorable influences that are difficult to shake. Excessive use of social media has the potential to cause addiction ^[8], which as a result could cause young people to spend more time on these platforms than they spend with their family and friends ^[9]. Their general health and social interactions may suffer as a result of this addiction. Additionally, social media witnesses the growing problem of the increase in online abuse and cyberbullying, which can have a negative impact on a person’s self-esteem and even violate their privacy ^[10]. The spread of misinformation and hatred online has also contributed to an uptick in violent crimes in society ^[11]. Receiving messages with vulgar language is a startling realization of this unwelcome and damaging phenomenon. The likelihood of encountering such vulgar remarks rises as social media use increases.

Vulgarity or obscenity with regards to language refers to terms used to describe the use of vulgar language, such as swearing, taboo words, or offensive expressions ^[12][13]. Unfortunately, such language has become more and more common in contemporary culture ^[14], especially on social media sites like Twitter ^[15]. The majority of the time, however, vulgar language is used in the context of online harassment and negativity. While there are some instances where vulgar language may be used in a positive context to convey informality, express anger, or establish a sense of belonging with a particular group ^[16], these are usually rare in comparison to its use in negative contexts or in closed message groups. Therefore, detecting such use of language quickly and effectively is necessary to allow social media platforms to efficiently moderate their contents.

2. Automatic Vulgar Word Extraction Method with Application to Vulgar Remark Detection in Chittagonian Dialect of Bangla

Table 1 shows research in Bengali on topics related to detecting vulgarity. Traditionally, vulgar expression lexicons have been developed as a means of vulgarity detection ^[17]. These lexicon-based approaches need to be updated frequently to remain effective, however. In contrast, machine learning (ML) techniques provide a more dynamic approach by classifying new expressions as either vulgar or non-vulgar without relying on predetermined lexicons. Deep learning has made significant contributions to the field of signal and image processing ^[18], diagnosis ^[19], wind forecasting ^[20] and time series forecasting ^[21]. Beyond lexicon-based techniques, vulgarity detection has been the subject of several studies. Moreover, numerous linguistic and psychological studies ^[22] have been carried out to comprehend the pragmatic applications ^[13] and various vulgar language forms ^[23]. For machine learning-related studies, for example, Eshan et al. ^[24] ran an experiment in which they classified data obtained by scraping the Facebook pages of well-known celebrities using the traditional machine learning classifiers multinomial naive Bayes, random forest, and SVM (support vector machine). They gathered unigram, bigram, and trigram features and weighted them using TF-IDF vectorizers. On datasets of various sizes, containing 500, 1000, 1500, 2000, and 2500 samples. The results showed that when using unigram features, a sigmoid kernel had the worst accuracy performance, and SVM with a linear kernel had the best accuracy performance. However, MNB demonstrated the highest level of accuracy for bigram and trigram features. In conclusion, TfidfVectorizer features outperformed CountVectorizer features when combined with an SVM linear kernel. Akhter et al. ^[25] suggested using user data and machine learning techniques to identify instances of cyberbullying in Bangla. They used a variety of classification algorithms, such as naive Bayes (NB), J48 decision trees, support vector machine (SVM), and k-nearest neighbors (KNN). A 10-fold cross-validation was used to assess how well each method performed. The results showed that SVM performed better than the other algorithms when it came to analyzing Bangla text, displaying the highest accuracy score of 0.9727. Holgate et al. ^[16] introduced a dataset of 7800 tweets from users whose demographics were known. Each instance of vulgar language use was assigned to one of six different categories by the researchers. These classifications included instances of aggression, emotion, emphasis, group identity signaling, auxiliary usage, and non-vulgar situations. They sought to investigate the practical implications of vulgarity and its connections to societal problems through a thorough analysis of this dataset. Holgate et al. obtained a macro F1 score of 0.674 across the six different classes by thoroughly analyzing the data that were gathered. Emon et al. ^[26] created a tool to find abusive Bengali text. They used various deep learning and machine learning-based algorithms to achieve this. A total of 4700 comments from websites like Facebook, YouTube, and Prothom Alo were collected in a dataset. These comments were carefully labeled into seven different categories. Emon et al. experimented with various algorithms to find the best one. The recurrent neural network (RNN) algorithm demonstrated the highest accuracy among the investigated methods, achieving a satisfying score of 0.82. Awal et al. ^[27] demonstrated a naive Bayes system made to look for abusive comments. They gathered a dataset of 2665 English comments from YouTube in order to evaluate their system. They then translated these English remarks into Bengali utilizing two techniques: (i) Bengali translation directly; (ii) Bengali translation using dictionaries. Awal et al. evaluated the performance of their system after the translations. Their system impressively achieved the highest accuracy of 0.8057, demonstrating its potency in identifying abusive content in the context of the Bengali language. Hussain et al. ^[28] suggested a method that makes use of a root-level algorithm and unigram string features to identify abusive Bangla comments. They gathered 300 comments for their dataset from a variety of websites, including Facebook pages, news websites, and YouTube. The dataset was split into three subsets, each of which contained 100, 200, and 300 comments. These subsets were used to test their system, which resulted in an average accuracy score of 0.689. Das et al. ^[29] carried out a study on detecting hate speech in Bengali and Romanized Bengali. They extracted samples from Twitter in order to gather the necessary information, producing a dataset with 5071 samples in Bengali and Romanized Bengali. They used a variety of training models in their study, including XML-RoBERTa, MuRIL, m-BERT, and IndicBERT. Following testing, they discovered that XML-RoBERTa had the highest accuracy, at 0.796. Sazzed ^[30] collected 7245 YouTube reviews manually and divided them into two categories: vulgar and non-vulgar. The purpose of this process was to produce two benchmark corpora for assessing vulgarity detection algorithms. Following the testing of several methods, the bidirectional long short-term memory (BiLSTM) model showed the most promising results, achieving the highest recall scores for identifying vulgar content in both datasets. Jahan et al. ^[31] created a dataset by using online comment scraping tools to collect comments from public Facebook pages, such as news and celebrity pages. SVM, random Forest, and AdaBoost were the three machine learning techniques used to categorize the comments for the detection of abusive content. Their approach, which was based on the random forest classifier, outperformed other methods in terms of accuracy and precision, scoring 0.7214 and 0.8007, respectively. AdaBoost, on the other hand, demonstrated the best recall performance, earning a score of 0.8131. Ishmam et al. ^[32] collected a dataset sourced from Facebook, categorized into six distinct classes. The dataset was enriched with linguistic and quantitative features, and the researchers employed a range of text preprocessing techniques, including punctuation removal, elimination of bad characters, handling hashtags, URLs, and mentions, as well as tokenization and stemming. They utilized neural networks, specifically GRUs (gated recurrent units), alongside other machine learning classifiers, to conduct classification tasks based on the historical, religious, cultural, social, and political contexts of the data. Karim et al. ^[33] used a combination of machine learning classifiers and deep neural networks to detect hate speech in Bengali. They analyzed datasets containing comments from Facebook, YouTube, and newspaper websites using a variety of models, including logistic regression, SVM, CNN, and Bi-LSTM. The researchers divided hate speech into four distinct categories: political, religious, personal, and geopolitical. With F1 scores of 0.78 for political hate speech, 0.91 for personal hate speech, 0.89 for geopolitical hate speech, and 0.84 for religious hate speech detection in the Bengali language, their results showed satisfying performance. Sazzed ^[34] created a transliterated corpus of 3000 comments from Bengali, 1500 of which were abusive and 1500 of which were not. As a starting point, they used a variety of supervised machine learning methods, such as deep learning-based bidirectional long short-term memory networks (BiLSTM), support vector machines (SVM), logistic regression (LR), and random forest (RF). The SVM classifier displayed the most encouraging results (with an F1 score of 0.827 ± 0.010) in accurately detecting abusive content. User comments from publicly viewable Facebook posts made by athletes, officials, and celebrities were analyzed in a study by Ahmed et al. ^[35]. The researchers distinguished between Bengali-only comments and those written in English or a mix of English and other languages. Their research showed that 14,051 initial comments in total, or approximately 31.9% of them, were directed at male victims. However, a significant number of the 29,950 comments, or 68.1% of the total, were directed at female victims. The study also highlighted how comments were distributed according to the different types of victims. A total of 9375 comments were directed at individuals who are social influencers. Among these, 5.98% (equivalent to 2633 comments) were aimed at politicians, while 4.68% (or 2061 comments) were focused on athletes. Additionally, 6.78% (about 2981 comments) of the comments were centered around singers, and the majority, which is 61.25% (totaling 26,951 comments), were directed at actors. For the classification of hate speech in the Bengali language, Romim et al. ^[36] used neural networks, including LSTM (long short-term memory) and BiLSTM (bidirectional LSTM). They used word embeddings that had already been trained using well-known algorithms such as FastText, Word2Vec, and Glove. The largest dataset of its kind to date, the extensive Bengali dataset they introduced for the research includes 30,000 user comments. The researchers thoroughly compared different deep learning models and word embedding combinations. The outcomes were encouraging as all of the deep learning models performed well in the classification of hate speech. However, the support vector machine (SVM) outperformed the others with an accuracy of 0.875. Islam et. al. ^[37] used large amounts of data gathered from Facebook and YouTube to identify abusive comments. To produce the best results, they used a variety of machine learning algorithms, such as multinomial naive Bayes (MNB), multilayer perceptron (MLP), support vector machines (SVM), decision tree, random forest, and SVM with stochastic gradient descent-based optimization (SGD), ridge classifier, and k-nearest neighbors (k-NN). They used a Bengali stemmer for preprocessing and random undersampling of the dominant class before processing the dataset. The outcomes demonstrated that, when applied to the entire dataset, SVM had the highest accuracy of 0.88. In their study, Aurpa et al. ^[38] used transformer-based deep neural network models, like BERT ^[39] and ELECTRA ^[40], to categorize abusive comments on Facebook. For testing and training, they used a dataset with 44,001 Facebook comments. The test accuracy for their models, which was 0.85 for the BERT classifier and 0.8492 for the ELECTRA classifier, showed that they were successful in identifying offensive content on the social media platform. Based on the comprehensive analysis of papers related to vulgarity detection and related topics like abusive and bullying detection, as well as detection in the low-resource language Bengali, several critical research gaps emerge. These gaps include the absence of a clear problem definition in some papers, the prevalence of small-sized datasets without a well-defined annotation process, and the lack of benchmarking efforts to assess dataset quality. Additionally, class imbalance in datasets remains an issue, and limited attention has been given to vulgarity detection in low-resource language Bengali, with only a single work ^[30] addressing this area. Many papers fail to specify the source of their datasets and conduct limited experiments. Field surveys are often superficial or nonexistent. Furthermore, none of the papers considered ethical considerations in data collection, such as preserving user privacy through dataset anonymization. Addressing these research gaps is essential for advancing the field of vulgarity detection and related areas, ensuring the development of more robust, ethical, and well-defined detection systems.