Hindi/Marathi-BART: An Efficient Sequence-to-Sequence Model for Abstractive Text Summarization for Hindi and Marathi language

Abstract

The efficacy of neural sequence-to-sequence models for text summarization has been dramatically enhanced by attention-based architectures. Despite the fact that these models are effective at summarizing English documents, they are not readily transferable to other languages, leaving room for development. This paper presents Hindi/Marathi-BART, a sequence-to-sequence model designed particularly for the Hindi and Marathi languages based on the BART architecture. The model is pre-trained on a large corpus of Hindi and Marathi texts to acquire language-specific features and then fine-tuned for abstractive summarization using benchmark datasets. Despite having substantially fewer parameters, Hindi/Marathi-BART outperforms other cutting-edge models in terms of ROUGE scores, as demonstrated by experimental results. Although the ROUGE score is commonly used to evaluate the quality of automatically generated summaries, it does not adequately convey the semantic similarity between automatically generated and reference summaries. To circumvent this limitation, we augment ROUGE with BERTScore, a more sophisticated evaluator that measures the token-level semantic similarity between the generated and reference summaries. Using Hindi/Marathi-BART can facilitate the development of natural language processing (NLP) applications for these two languages. We will disseminate the model to the research community in an effort to stimulate additional investigation and application.