Graph Based Extractive News Articles Summarization Approach leveraging Static Word Embeddings

Abstract

With enormous and voluminous data being generated on a regular basis at an exponential speed, there is a demanding need for concise and relevant information to be available for the masses. Traditionally, lengthy textual contents are manually summarized by Linguists or Domain Experts, which are highly time consuming and unfairly biased. There is a dire need for Automatic Text Summarization approaches to be introduced in this broad spectrum. Extractive Summarization is one such approach where the salient information or excerpts are identified from a source and extracted to generate a concise summary. TextRank is an unsupervised extractive summarization technique incorporating graph-based ranking of extracted texts and finding the most relevant excerpts to generate a concise summary. In this paper, the prospects of a domain agnostic algorithm like TextRank for various domains of News Article Summarization are explored, exploring its efficiency in domain specific tasks and conveniently drawing various insights. NLP based pre-processing approaches and Static Word Embeddings were leveraged with semantic cosine similarity for the efficient ranking of textual data and performance evaluation on various domains of BBC News Articles Summarization datasets through ROUGE metrics. A commendable ROUGE score is achieved.