Enterprise AI Analysis

Robust Preprocessing and Feature Engineering Pipeline

The proposed Urdu Toxicity Detection Model (UTDM) framework employs a rigorous methodology starting with comprehensive data preprocessing. This involves removing other language data (English, Arabic, numerical digits), punctuation marks, and repetitive characters, followed by tokenization, stemming, and data normalization. A custom Urdu stop words list further refines the text. Feature engineering is then applied using Term Frequency-Inverse Document Frequency (TF-IDF), Bag-of-Words (BoW), and N-gram techniques to capture both term presence/frequency and contextual patterns, crucial for Urdu's complex morphology. To address class imbalance, the Synthetic Minority Over-sampling Technique (SMOTE) is utilized, creating synthetic examples for minority classes and ensuring balanced data distribution for training. This multi-stage approach ensures data quality and prepares the text for effective classification by both machine learning and deep learning models.

Random Forest Achieves Benchmark Performance

Experimental results demonstrate that the Random Forest (RF) classifier significantly outperforms all other evaluated Machine Learning (ML) and Deep Learning (DL) models in detecting Urdu toxicity. RF achieved an exceptional precision of 0.97, recall of 0.99, F1-score of 0.98, and overall accuracy of 0.99 (99%). This superior performance is attributed to RF's robustness on smaller-to-medium sized datasets and its effectiveness in exploiting features generated by TF-IDF, BoW, and N-grams. While DL models (LSTM, BiLSTM, GRU) showed potential, especially GRU with better generalization for minority classes, they did not match RF's performance or computational efficiency. The 5-fold cross-validation further validated RF's stable and reliable generalization capacity across all five toxicity classes.

ML Model F1-Scores (5-fold cross-validation)

DL Model F1-Scores (5-fold cross-validation)

Advancing NLP for Low-Resource Languages

This study provides a robust foundation for advancing Urdu natural language processing by creating the first comprehensive multi-label, multi-class Urdu toxicity dataset and proposing the highly effective UTDM framework. While Random Forest demonstrated superior performance in this work due to dataset characteristics and feature engineering, the study acknowledges limitations such as potential annotation bias and the current scope of the dataset. Future research is vital, focusing on expanding the dataset with more diverse Urdu texts, exploring semi-supervised or unsupervised learning to reduce reliance on manual labeling, and developing advanced pre-trained transformer models specifically fine-tuned for Urdu. Furthermore, integrating human-in-the-loop moderation and explainable AI techniques (e.g., SHAP, attention maps) will be crucial to mitigate risks of censorship and bias, ensuring a safer and more inclusive digital environment for Urdu-speaking communities.