BanglaHealth: A Bengali paraphrase dataset on health domain

In the landscape of natural language processing (NLP) research, the availability of comprehensive datasets plays a pivotal role in advancing various tasks, including paraphrasing. However, for languages such as Bengali, the availability of such datasets remains limited, particularly in specialized domains like health. Recognizing this gap, this study endeavours to address the scarcity of resources by presenting a novel Bengali paraphrasing dataset specifically tailored to the health domain. The dataset construction process involved sourcing sentences from Bengali newspapers, focusing on health-related content. Due to the dearth of existing datasets in Bengali and the specialized nature of paraphrasing, particularly in the health domain, this endeavour necessitated the development of a unique methodology. This methodology included the development of a script for data extraction, pre-processing, translation of Bengali sentences into English, paraphrasing of the English sentences, and subsequent translation back into Bengali to generate paraphrased versions of the original sentences. A user study involving 100 participants was conducted to evaluate 500 sample Bengali paraphrased sentences to identify the most suitable library (<i>VamsiT5 Paws</i>) for creating the paraphrasing dataset. As such, a total of 200,000 sentences were extracted and paraphrased in this process. This dataset holds significant potential to foster advancements in paraphrasing research that facilitate the development of language models; and ultimately contribute to the broader goal of enhancing NLP capabilities in Bengali, particularly in specialized domains like health.