Unveiling the Effectiveness of NLP-Based DL Methods for Urdu Text Analysis

The analysis of text data has become a significant challenge while its size is gradually increasing in massive amounts. Various textual analysis methods exist, dealing with different processing styles due to multiple data types, mainly for English. Therefore, the other low-resource languages are difficult to process due to the unavailability of intelligent methods. Similarly, Urdu, as a low-resource language, requires effective methods based on machine learning or deep learning mechanisms. Our study has identified the rarely used pure Urdu text dataset, an effective combination of embeddings, and the best combination of hyperparameters for DL methods trained on that dataset. According to the evaluation results, our study has also determined the best methods regarding embeddings, hyperparameters, and overall performance. Moreover, combining pre-trained BERT embeddings with the fine-tuned BiLSTM and BERT was the best method to cope with Urdu as a low-resource language. As per the findings, our study recommends the pre-trained embedding models and hyperparameters settings for Urdu text classification analysis.