Extractive Summarization of Urdu Language using Deep Learning Techniques on a Custom Dataset

Urdu, a major language in South Asia, has been overlooked in the field of Natural Language Processing (NLP), despite its widespread and popular use. Text summarization, a fundamental part of NLP, allows for the processing of large amounts of textual content into short summaries. Both extractive and abstractive summary approaches are widely used; however, extractive summarization in Urdu has gotten less attention. This research study fills that gap by offering a new, customized Urdu corpus comprised of different articles from multiple fields along with their extractive summaries and topics. Deep learning techniques are used in this research, including Bidirectional Long Short-Term Memory (LSTM) networks and a customized (BERT) model by adding extra attention and dense layers, that were trained on this corpus. The efficiency of these models is assessed using ROUGE metrics, with the LSTM model obtaining ROUGE-1, ROUGE-2, and ROUGE-L scores of 0.48, 0.23, and 0.48 respectively. In comparison, the BERT model has ROUGE scores of 0.65, 0.18, and 0.65 for ROUGE-1, ROUGE-2, and ROUGE-L, respectively. Despite these results, the BERT model performs well on unseen data, producing precise summaries. This study emphasizes the significance of creating language-specific NLP tools and shows the potential for deep learning applications in under-represented languages.