Automatic Gujarati Text Summarization Using Natural Language Processing: A Gujarati-Specific Abstractive Framework

Abstract

With the ever-increasing digital information, there is a high demand for automatic text summarization systems to produce meaningful summaries from longer documents. Despite significant efforts in automatic text summarization for high resource languages like English, research on automatic text summarization in Gujarati is limited because of the lack of linguistic resources, a lack of annotated datasets, and the difficult grammar of the Gujarati language. Previous multilingual transformer models like mBART, IndicBART and mT5 have shown promising results; however, they tend to produce grammatically incorrect, repetitive and incongruent summaries for Gujarati documents. In this study, we present a Gujarati specific abstractive text summarization system that leverages an improved multilingual mT5 model with a linguistic preprocessing step on the input text, morphological normalization, named entity preservation and coverage-aware decoding step. The proposed system takes as input any paragraph or article of Gujarati text or any document of large text, and produces one or two sentence summaries which are semantically equivalent to the original document and contain the information. The proposed framework would achieve better quality of summarization, grammatical correctness, semantic consistency, computational efficiency, and solve the issues with low-resource languages from India. To prove the superiority of the proposed model over the existing multilingual summarization models, automatic evaluation metrics such as ROUGE, BLEU, BERTScore will be employed in addition to human evaluation.