Bidirectional Encoder Approach for Abstractive Text Summarization of Urdu Language

Abstract

The fast pace of accumulating textual data in the online sphere has made it laborious to get out handy information from a profuse amount of information. NLP's area: automate text summarization, yields a great quality and considerable gist; abstracts and summaries of written texts of myriad human languages. Several attempts have been carried out previously in extractive summarization systems; however, research in abstractive summarization in the Urdu language has not been studied well so far. Urdu is a very rich language in terms of literary sources and it requires serious research efforts to generate abstractive summaries. In this research, we employ a composition of abstractive and extractive algorithms in an automated text summarization system for the Urdu language. In extractive summaries, we use word frequency, Sentence weight, and TF-IDF algorithms. Further, a hybrid method is introduced to improve the results of extractive summaries. Bidirectional Encoder Representations from Transformers (BERT) model is used to process the summaries generated by hybrid method for generation of abstractive summary. To evaluate the system-generated summaries, the assistance of the experts of Urdu language is reaped.