Implementation of Bangla Extractive Update Summarization Task on BUSUM-BNLP-Dataset: A Multi-Document Update Summarization Corpus

The research inspires automated multi-document update summarization to perform a more complete linguistic (semantic) assessment of Bangla newspaper articles, research articles, or multiple texts by selecting the most essential ideas. Automated update summarizer methods are critical in identifying relevant information and up-to-date material for the user, minimizing the amount of time and effort necessary to read an entire article or many articles on a similar topic. Many updated summary approaches have been developed and implemented in the English language, while updated summarization in Bengali remains restricted. Keeping this in mind, we want to devise a straightforward method for developing a Bengali automatic updated summary task. To collaborate on this research, we proposed BUSUM-BNLP, a multi-document updated summarization Bangla dataset that is divided into six categories: international, sports, politics, entertainment, Bangladesh, and business. Each article has been divided into two groups, A and B. Essentially, every pair of A and B has five news article files, each with a title, date-time, and pertinent news content. Similarly, for each set of documents, we developed three human-written summaries. To generate a quality dataset, we reviewed 1,000 Bangla news articles from major newspaper e-platforms (Prothom Alo, Jugantor, Kaler Kantho, BBC Bangla) and researched certain Bangla language and summarizing norms. Since there are two kinds of summarization methods: extractive and abstractive. In this paper, we created a TF-IDF-based model and used a pre-trained SentenceRank to obtain an updated Bengali summarization model. We tested these two models using our BUSUM-BNLP dataset, and the experienced TF-IDF model outperformed the pre-trained SenteneRank model.