Abstract:

Automatic text summarization is increasingly important for managing large volumes of information, as it allows extracting the essential content of documents in a concise form. Neural sequence-to-sequence models with attention mechanisms have become a leading approach for abstractive summarization, as attention focuses the model on the most relevant parts of the input text. In this work, we compare two prominent attention mechanisms – the additive Bahdanau attention and the dot-product Luong attention – within a recurrent sequence-to-sequence (RNN Seq2Seq) model for text summarization. We conduct experiments on a news article dataset to evaluate how each attention type affects summary quality and model efficiency. The results show measurable differences: the Luong (dot-product) attention model achieved slightly higher ROUGE
scores and better generalization than the Bahdanau model in our tests. Qualitative analysis of generated summaries indicates that while both models capture the main ideas, the Bahdanau attention variant tended to miss some details or repeat content, whereas the Luong variant produced more coherent and concise summaries. These findings highlight the impact of attention mechanism choice on summarization performance and provide insights for designing efficient summarization systems.