Abstractive Text Summarization for the Urdu Language: Data and Methods

The task of abstractive text summarization aims to automatically generate a short and concise summary of a given source article. In recent years, automatic abstractive text summarization has attracted the attention of researchers because large volumes of digital text are readily available in multiple languages on a wide range of topics. Automatically generating precise summaries from large text has potential application in the generation of news headlines, a summary of research articles, the moral of the stories, media marketing, search engine optimization, financial research, social media marketing, question-answering systems, and chatbots. In literature, the problem of abstractive text summarization has been mainly investigated for English and some other languages. However, it has not been thoroughly explored for the Urdu language despite having a huge amount of data available in digital format. To fulfill this gap, this paper presents a large benchmark corpus of 2,067,784 Urdu news articles for the Urdu abstractive text summarization task. As a secondary contribution, we applied a range of deep learning (LSTM, Bi-LSTM, LSTM with attention, GRU, Bi-GRU, and GRU with attention), and large language models (BART and GPT-3.5) on our proposed corpus. Our extensive evaluation on 20,000 test instances showed that GRU with attention model outperforms the other models with ROUGE-
$1=46.7$
, ROUGE-
$2=24.1$
, and ROUGE-L =48.7. To foster research in Urdu, our proposed corpus is publically and freely available for research purposes under the Creative Common Licence.