An Advanced Text-Rank Model For Extractive Text Summarization Using Semantic Similarity And Redundancy Control

Abstract

Extractive text summarization is considered a significant issue within the field of natural language processing, especially in the context of ever-increasing volumes of digital data. While graph-based techniques like TextRank have proven effective at determining salient sentences, there have been issues with their inability to incorporate semantic relations and avoid redundant summaries. In this paper, we will propose a TextRank model that utilizes semantic similarity based on pre-trained transformers and employs Maximal Marginal Relevance (MMR) technique to reduce redundancy in the output. The new algorithm proposed in this work builds on the classic TextRank by adding the following: (1) replacing cosine similarity between TF-IDF vectors with more sophisticated sentence embeddings from the BERT architecture, (2) employing MMR for selecting sentences, and (3) introducing additional features of sentences’ positions and lengths. We tested the proposed method on two standard corpora: CNN/Daily Mail and DUC 2002. The experimental results have shown that the proposed advanced version of the TextRank algorithm is able to provide significant improvements compared to the baseline TextRank algorithm. The ROUGE-1 score was increased by 18.3%, the ROUGE-2 score was increased by 17.9%, and the ROUGE-L score was increased by 19.9%. Our model is computationally efficient and provides more coherent summaries.