Turkish abstractive text summarization using pretrained sequence-to-sequence models

Abstract The tremendous amount of increase in the number documents available on Web has turned finding relevant piece information into a challenging, tedious, and time-consuming activity. Accordingly, automatic text summarization become an important field study by gaining significant attention from researchers. Lately, with advances deep learning, neural abstractive sequence-to-sequence (Seq2Seq) models gained popularity. There have been many improvements these such as use pretrained language (e.g., GPT, BERT, XLM) Seq2Seq BART T5). These addressed certain shortcomings improved upon challenges saliency, fluency, semantics which enable generating higher quality summaries. Unfortunately, research attempts were mostly limited to English language. Monolingual BERT multilingual released recently providing opportunity utilize state-of-the-art low-resource languages Turkish. In this study, we make obtain results two large-scale Turkish datasets, TR-News MLSum, for task. Then, title datasets establish hard baselines generation task both datasets. We show that input substantial importance success tasks. Additionally, provide extensive analysis including cross-dataset evaluations, various options, effect preprocessing ROUGE evaluations It is shown monolingual outperform all tasks across Lastly, qualitative generated summaries titles are provided.