An Efficient Deep Learning Approach for Extractive Arabic Text Summarization Based on Multiple Encoders and a Single Decoder

Abstract

The amount of textual content produced has grown dramatically over the last few thousand years, creating a wealth of information for analysis and knowledge extraction. Daily textual data’s increasing volume and complexity, including social media posts, news articles, emails, and text messages, make consuming and processing all the information difficult. This also means manually sifting through large volumes of text to find relevant information can be time-consuming and difficult. Text summarizing addresses this challenge by automatically condensing text into a more concise format so that users can quickly and easily access the most critical information. In today’s data-driven environment, it has become an essential tool used in multiple contexts, including text analysis, content-based recommendations, and information retrieval. The paper presents a novel extractive Arabic text summarization approach in a single-document framework. The proposed method utilizes a Modified Sequence-to-Sequence (MSTS) algorithm, which incorporates multi-layer encoders and a single-layer decoder into the sequence-to-sequence model. A new Arabic benchmark dataset called HASD is also introduced, which contains 43k articles with their extractive and abstractive summaries. To assess the accuracy of the generated summaries, the authors introduce a new evaluation measure called eval_summary. This measure aids in identifying the most accurate summary from the multiple summarizations produced by the model. The proposed model is tested on the proposed HASD and EASC benchmarks and evaluated using Rouge, Bleu. The experimental results show satisfactory results compared to state-of-the-art methods.