Nisar Hussain

This study addresses the task of detecting offensive language in Urdu text on social media, where harmful and offensive comments can lead to emotional distress. Using a dataset of Urdu comments collected from YouTube news channels, both deep learning (DL) and machine learning (ML) models were employed for offensive language detection. Feature extraction was performed using Term Frequency–Inverse Document Frequency (TF-IDF) and the Count Vectorizer approach to capture unigrams, bigrams, and trigrams. Four ML models—Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and Gaussian Naive Bayes (NB)—were evaluated, with SVM achieving comparatively stronger performance among the ML approaches. For the deep learning framework, Bidirectional Long Short-Term Memory (Bi-LSTM) and Convolutional Neural Network (CNN) architectures were implemented using custom Word2Vec and FastText embeddings. The experimental results indicated that the CNN model outperformed the other evaluated approaches, suggesting its effectiveness for detecting offensive language in Urdu text. Smart citations: https://scite.ai/reports/10.61467/2007.1558.2026.v17i1.1240 Dimensions.Open Alex.

In this work, we investigate sentiment classification on Twitter using the Sentiment140 dataset and compare traditional deep learning approaches, including CNN, BiLSTM, and GRU, with a lightweight transfer learning model, DistilBERT. The results indicate that the CNN, BiLSTM, and GRU models achieve accuracy levels ranging from 71% to 79%, whereas the fine-tuned DistilBERT model attains an accuracy of 86% with an F1-score of 0.86. These findings suggest that, while conventional neural models tend to focus on more local patterns, lightweight transformer-based architectures are better able to capture[HEDGING-END] broader linguistic representations, while still maintaining relatively efficient computational performance. Overall, this study offers a systematic and practical comparison of sentiment classification models, supported by a clearly defined methodology, empirical results, and corresponding tables and figures. The analysis may provide useful guidance for researchers and practitioners interested in sentiment analysis applications for social media data. Smart citations: https://scite.ai/reports/10.61467/2007.1558.2026.v17i1.1241 Dimensions.Open Alex.

Depression is now one of the most common mental health concerns in the digital era, calling for powerful computational tools for its detection and its level of severity estimation. A multi-level depression severity detection framework in the Reddit social media network is proposed in this study, and posts are classified into four levels: minimum, mild, moderate, and severe. We take a dual approach using classical machine learning (ML) algorithms and recent Transformer-based architectures. For the ML track, we build ten classifiers, including Logistic Regression, SVM, Naive Bayes, Random Forest, XGBoost, Gradient Boosting, K-NN, Decision Tree, AdaBoost, and Extra Trees, with two recently proposed embedding methods, Word2Vec and GloVe embeddings, and we fine-tune them for mental health text classification. Of these, XGBoost yields the highest F1-score of 94.01 using GloVe embeddings. For the deep learning track, we fine-tune ten Transformer models, covering BERT, RoBERTa, XLM-RoBERTa, MentalBERT, BioBERT, RoBERTa-large, DistilBERT, DeBERTa, Longformer, and ALBERT. The highest performance was achieved by the MentalBERT model, with an F1-score of 97.31, followed by RoBERTa (96.27) and RoBERTa-large (96.14). Our results demonstrate that, to the best of the authors’ knowledge, domain-transferred Transformers outperform non-Transformer-based ML methods in capturing subtle linguistic cues indicative of different levels of depression, thereby highlighting their potential for fine-grained mental health monitoring in online settings.

The detection of abusive language in Roman Urdu is important for secure digital interaction. This work investigates machine learning (ML), deep learning (DL), and transformer-based methods for detecting offensive language in Roman Urdu comments collected from YouTube news channels. Extracted features use TF-IDF and Count Vectorizer for unigrams, bigrams, and trigrams. Of all the ML models—Random Forest (RF), Logistic Regression (LR), Support Vector Machine (SVM), and Naïve Bayes (NB)—the best performance was achieved by the same SVM. DL models involved evaluating Bi-LSTM and CNN models, where the CNN model outperformed the others. Moreover, transformer variants such as LLaMA 2 and ModernBERT (MBERT) were instantiated and fine-tuned with LoRA (Low-Rank Adaptation) for better efficiency. LoRA has been tuned for large language models (LLMs), a family of advanced machine learning frameworks, based on the principle of making the process efficient with extremely low computational cost with better enhancement. According to the experimental results, LLaMA 2 with LoRA attained the highest F1-score of 96.58%, greatly exceeding the performance of other approaches. To elaborate, LoRA-optimized transformers perform well in capturing detailed subtleties of linguistic nuances, lending themselves well to Roman Urdu offensive language detection. The study compares the performance of conventional and contemporary NLP methods, highlighting the relevance of effective fine-tuning methods. Our findings pave the way for scalable and accurate automated moderation systems for online platforms supporting multiple languages.

Hate speech is a complex topic that transcends language, culture, and even social spheres. Recently, the spread of hate speech on social media sites like Facebook has added a new layer of complexity to the issue of online safety and content moderation. This study seeks to minimize this problem by developing an Arabic script-based tool for automatically detecting hate speech in Roman Urdu, an informal script used most commonly for South Asian digital communications. Roman Urdu is relatively complex as there are no standardized spellings, leading to syntactic variations, which increases the difficulty of hate speech detection. To tackle this problem, we adopt a holistic strategy using a combination of six machine learning (ML) and four Deep Learning (DL) models, a dataset from Facebook comments, which was preprocessed (tokenization, stopwords removal, etc.), and text vectorization (TF-IDF, word embeddings). The ML algorithms used in this study are LR, SVM, RF, NB, KNN, and GBM. We also use deep learning architectures like CNN, RNN, LSTM, and GRU to increase the accuracy of the classification further. It is proven by the experimental results that deep learning models outperform the traditional ML approaches by a significant margin, with CNN and LSTM achieving accuracies of 95.1% and 96.2%, respectively. As far as we are aware, this is the first work that investigates QLoRA for fine-tuning large models for the task of offensive language detection in Roman Urdu.

Depression is now one of the most common mental health concerns in the digital era, calling for powerful computational tools for its detection and its level of severity estimation. A multi-level depression severity detection framework in the Reddit social media network is proposed in this study, and posts are classified into four levels: minimum, mild, moderate, and severe. We take a dual approach using classical Machine Learning (ML) algorithms and recent Transformer-based architectures. For the ML track, we build ten classifiers, including Logistic Regression, SVM, Naive Bayes, Random Forest, XGBoost, Gradient Boosting, K-NN, Decision Tree, AdaBoost, and Extra Trees, with two recently proposed embedding methods, Word2Vec and GloVe embeddings, and we fine-tune them for mental health text classification. Of these, XGBoost yields the highest F1-score of 94.01 using GloVe embeddings. For the deep learning track, we fine-tune ten Transformer models, covering BERT, RoBERTa, XLM-RoBERTa, MentalBERT, BioBERT, RoBERTa-large, DistilBERT, DeBERTa, Longformer, and ALBERT. The highest performance was achieved by the MentalBERT model with an F1-score of 97.31, followed by RoBERTa (96.27) and RoBERTa-large (96.14). Our results demonstrate that, to the best of the authors’ knowledge, domain-transferred Transformers outperform non-Transformer-based ML methods in capturing subtle linguistic cues indicative of different levels of depression, thereby highlighting their potential for fine-grained mental health monitoring in online settings.

With the rapid expansion of social media, detecting offensive language has become critically important for healthy online interactions. This poses a considerable challenge for low-resource languages such as Roman Urdu which are widely spoken on platforms like Facebook. In this paper, we perform a comprehensive study of offensive language detection models on Roman Urdu datasets using both Machine Learning (ML) and Deep Learning (DL) approaches. We present a dataset of 89,968 Facebook comments and extensive preprocessing techniques such as TF-IDF features, Word2Vec, and fastText embeddings to address linguistic idiosyncrasies and code-mixed aspects of Roman Urdu. Among the ML models, a linear kernel Support Vector Machine (SVM) model scored the best performance, with an F1 score of 94.76, followed by SVM models with radial and polynomial kernels. Even the use of BoW uni-gram features with naive Bayes produced competitive results, with an F1 score of 94.26. The DL models performed well, with Bi-LSTM returning an F1 score of 98.00 with Word2Vec embeddings and fastText-based Bi-RNN performing at 97.00, showcasing the inference of contextual embeddings and soft similarity. The CNN model also gave a good result, with an F1 score of 96.00. The CNN model also achieved an F1 score of 96.00. This study presents hybrid ML and DL approaches to improve offensive language detection approaches for low-resource languages. This research opens up new doors to providing safer online environments for widespread Roman Urdu users.

Thisstudy introduces a new model for detecting insults in Roman Urdu, filling an important gap in natural language processing (NLP) for low-resource languages. The transliterated nature of Roman Urdu also poses specific challenges from a computational linguistics perspective, including non-standardized grammar, variation in spellings for the same word, and high levels of code-mixing with English, which together make automated insult detection for Roman Urdu a highly complex problem. To address these problems, we created a large-scale dataset with 46,045 labeled comments from social media websites such as Twitter, Facebook, and YouTube. This is the first dataset for insult detection for Roman Urdu that was created and annotated with insulting and non-insulting content. Advanced preprocessing methods such as text cleaning, text normalization, and tokenization are used in the study, as well as feature extraction using TF–IDF through unigram (Uni), bigram (Bi), trigram (Tri), and their unions: Uni+Bi+Trigram. We compared ten machine learning algorithms (logistic regression, support vector machines, random forest, gradient boosting, AdaBoost, and XGBoost) and three deep learning topologies (CNN, LSTM, and Bi-LSTM). Different models were compared, and ensemble ones were proven to give the highest F1-scores, reaching 97.79%, 97.78%, and 95.25%, respectively, for AdaBoost, decision tree, TF–IDF, and Uni+Bi+Trigram configurations. Deeper learning models also performed on par, with CNN achieving an F1-score of 97.01%. Overall, the results highlight the utility of n-gram features and the combination of robust classifiers in detecting insults. This study makes strides in improving NLP for Roman Urdu, yet further research has established the foundation of pre-trained transformers and hybrid approaches; this could overcome existing systems and platform limitations. This study has conscious implications, mainly on the construction of automated moderation tools to achieve safer online spaces, especially for South Asian social media websites.