Work

A Bengali-Hindi-Telugu parallel corpus for enhanced literary machine translation

Year: 2026

Type: data-paper

Abstract: Machine translation is an important application in natural language processing. It relies heavily on bilingual and multilingual corpora. India’s rich linguistic diversity has made it challenging to have large and good-quality parallel text corpus for Indian languages. This data article introduces BHT25, a carefully curated parallel corpus of approximately 27,149 trilingual sentence triplets in Bengali, Hindi, and Telugu. The corpus is compiled from emotion-rich literary works of renowned Indian authors such as Rabindranath Tagore and Sarat Chandra Chattopadhyay. Unlike existing datasets that primarily focus on news or government texts, BHT25 emphasizes literary and archaic language varieties, including traditional forms such as Bengali Sadhu Bhasha . The development process of BHT25 followed structured data cleaning, preprocessing, sentence alignment, and human validation procedures. Each sentence triplet is assigned a unique identifier to ensure traceability and a systematic analysis. The dataset is released under the CC-BY-4.0 license on Hugging Face, along with preprocessing scripts and documentation, to support research on Indic machine translation and multilingual literary processing. An extended version, annotated with four emotion labels (joy, sadness, anger, fear) and pairwise semantic similarity scores, is also included to support future research on emotion-aware translation and semantic consistency.

Source: Data in Brief

DOI: 10.1016/j.dib.2026.112775