Abstractive Text Summarization for Afaan Oromo using sequence-to-sequence RNNs

Abstract

Recently, the volume of textual data has rapidly increased, which has generated a valuable resource for extracting and analyzing information. This information must be summarized to retrieve useful knowledge within a reasonable time period. Text summarization is one of the main issues in natural language processing in recent years. Text summarization is a technique for generating concise and succinct summaries from long texts that focuses on the most relevant information while preserving the text's overall comprehensive meaning. In this thesis, we propose a method of generating short abstractive summaries for Afaan Oromo texts using some basic NLP techniques with sequence to-sequence recurrent neural network algorithms. The dataset has been collected from various Afaan Oromo online resources, including BBC Afaan Oromo news, Kallacha Oromiyaa newspapers, Fana Afaan Oromo news, Afaan Oromo Watchtower study text, Afaan Oromo Bible text, Ethiopian News Agency (ENA), Ethiopian Press Agency (EPA) and Jehovah Witnesses (JW) publication texts, etc. Then, the dataset has been preprocessed. Finally, the abstractive summary has been generated using sequence-to-sequence RNNs deep learning techniques. In order to evaluate the performance of the proposed system, we have used a well-known metric ROUGE for evaluating our model. The performance of four summarizers (E1, E2, E3 and E4) was measured using ROUGE-1. The average F1-Score values obtained were 0.16, 0.24, 0.26 and 0.34 respectively. Among them, E4 exhibited the highest performance, outperforming the other summarizers.