Enhancing Extractive Text Summarization of Long Documents Using Features Dimensionality Reduction

Automatic Text Summarization (ATS) task is considered one of the most challenging tasks in Natural Language Processing field, especially for long documents. There are many techniques for ATS, including Artificial Intelligence (AI) approaches, statistical-based approaches, graph-based approaches, etc., but the generated summaries are still less accurate than human summaries. This research presents an extractive Neural-based model to summarize long documents using both attentive global and local contexts, Word2Ve
$c$
embeddings, and a bidirectional Gated Recurrent Unit (GRU) as a Sentence Encoder. The proposed model uses a Multi-Layer Perceptron (MLP) that consists of two linear layers with Rectified Linear Activation Units (RELUs) to reduce features dimensionality. The proposed model is evaluated on PubMed and arXiv datasets using ROUGE metrics, and the evaluation results show remarkable and promising improvements for ROUGE-1, ROUGE-2, and ROUGE-L.