BERT Tokenization and Hybrid-Optimized Deep Recurrent Neural Network for Hindi Document Summarization

Abstract

Text summarization generates a concise summary of the available information by determining the most relevant and important sentences in the document. In this paper, an effective approach of document summarization is developed for generating summary of Hindi documents. The developed deep learning-based Hindi document summarization system comprises of a number of phases, such as input data acquisition, tokenization, feature extraction, score generation, and sentence extraction. Here, a deep recurrent neural network (Deep RNN) is employed for generating the scores of the sentences based on the significant features, wherein the weights and learning parameters of the deep RNN are updated by using the devised coot remora optimization (CRO) algorithm. Moreover, the developed CRO-Deep RNN is examined for its efficacy considering metrics, like recall-oriented understudy for gisting evaluation (ROUGE), recall, precision, and f-measure, and is found to have attained values of 80.896%, 95.700%, 95.051%, and 95.374%, respectively.