Unsupervised Extractive News Articles Summarization leveraging Statistical, Topic-Modelling and Graph-based Approaches

Abstract

Due to the presence of large amounts of data and its exponential level generation, the manual approach of summarization takes more time, is biased, and needs linguistic professional experts. To avoid these substantial issues or to generate a succinct summary report, automatic text summarization is very much important. Three different approaches namely the statistical approach such as Term Frequency Inverse Document Frequency(TF-IDF), the topic modeling approach such as Latent Semantic Analysis (LSA), and graph-based approaches such as TextRank were applied to generate a concise summary for the benchmark the British Broadcasting Corporation (BBC) news articles summarization dataset. The domain-specific implementations of each approach in the five domains of the dataset and domain-agnostic prospects were explored in the paper while drawing various insights. The generated summaries were evaluated using the Recall-Oriented Understudy for Gisting Evaluation (ROUGE) framework, leveraging precision, recall, and f-measure metrics. The approaches were not only able to achieve a commendable ROUGE score but also outperform the previous works on the dataset.