الفهرس | Only 14 pages are availabe for public view |
Abstract Automatic text summarization (ATS) for long documents is a very challenging task. A long document includes more than one topic, so it is required to generate a summary that covers the most important information of different topics in the input document. ATS has attracted scientific researchers since the 1950s, and there are many ATS techniques, but the generated summaries are much less accurate than the human summaries. Accordingly, the automatic summarization task is considered one of the most challenging tasks in Natural Language Processing, especially for long documents. ATS has different approaches like extractive, abstractive, and hybrid approaches. This thesis focuses on the extractive approach based on Neural Networks. It enhances an existing extractive Neural-based attentive model to summarize long documents using two different methods. The existing model uses a bidirectional Gated Recurrent Unit (GRU) as sentence encoder and the word embeddings averaging as word encoder. It also uses a Multi-Layer Perceptron (MLP) that consists of one linear layer with ReLU activation function for dimensionality reduction. The first proposed method is to replace the bidirectional GRU with two bidirectional GRUs as word and sentence encoders. Then, two ensembles are proposed by applying an average ensemble to the proposed model with two different Neural-based models, separately. The second proposed method is to replace the MLP with another that consists of two linear layers and ReLUs. This thesis uses ROUGE-1, ROUGE-2, and ROUGE-L metrics to evaluate the two proposed methods. The first method is evaluated on the PubMed dataset. The evaluation results of the first method on the PubMed dataset show promising improvements of 0.14% and 0.97% for ROUGE-1, 0.33% and 1.12% for ROUGE-2, and 0.25% for ROUGE-L. The second method is evaluated on the PubMed and arXiv datasets. The evaluation results of the second method on the PubMed dataset show promising improvements of 1.20 " ~ " 1.50% for ROUGE-1, 1.29 " ~ " 1.54% for ROUGE-2, and 1.13 " ~ " 1.33% for ROUGE-L. The evaluation results of the second method on the arXiv dataset show promising improvements of 0.10 " ~ " 1.87% for ROUGE-1, 0.10 " ~ " 1.84% for ROUGE-2, and 0.003 " ~ " 1.48% for ROUGE-L |