Abstractive Tigrigna Text Summarization using Deep Learning Approach

Abstract

Text summarization has become essential due to the vast amounts of text data shared online. It is the task of condensing long sequences of text into short, concise, and expressive summaries. Two basic methods exist: extractive and abstractive. In Tigrigna, some studies used extractive methods, focusing on selecting portions of text without addressing underlying meaning. We proposed abstractive text summarization for Tigrigna to generate semantics-based summaries. Abstractive summarization rephrases or reorganizes long text to produce semantically equivalent summaries, possibly with new words or phrases. Tigrigna has not been researched in this way, and the task is difficult since there is no structured dataset, pre-trained word embedding, or summarization model prepared for it. To address these challenges, we applied deep learning models. A dataset of 1,167 structured input paragraphs and reference summaries was prepared for training and evaluation. Different embedding methods, including FastText and Byte Pair Encoding, were trained on about 320 MB of data. To reduce the effect of noisy stopwords dominating embeddings, FastText was trained with stopword removal and down-sampling of frequent words. For Tigrigna abstractive summarization, two models (Seq2Seq LSTM and Transformer) were evaluated. The Seq2Seq works sequentially, whereas the Transformer operates in parallel. An attention mechanism was added to Seq2Seq, while Transformer uses self-attention. Among tested model–embedding matches, Seq2Seq with attention and FastText with down-sampling showed superior performance, achieving accuracy of 0.72 and ROUGE scores of R-1=0.20, R-2=0.183, and R-N=0.17. This work pioneers Tigrigna abstractive summarization, marking a foundational step for future research.