by Lukas Christian Nielsen & Sebastian Lindegaard Veile
Automatic text summarization is a useful tool to address this by enabling the reduction of longer texts down to their most important content. For English and other larger languages, automatic summarization is a relatively well studied field with a plethora of different summarization models. For lower resource languages like Danish, this is not the case. In this thesis, we explore the use of the popular BERT architecture for improving research in automatic summarization for Danish. Furthermore, we introduce TSAuBERT (Textual Similarity Assessment using BERT), a BERT based evaluation metric used for assessing the quality of automatically generated summaries, attempting to address the pitfalls of the current summarization evaluation protocol. We fine-tune 27 different BERT based summarization models using six Danish datasets, two different BERT models, and three summarization strategies. Our results provide new state of the art scores for each Danish datasets and solid scores for future research to improve upon. We further find that the existing Danish BERT model struggle compared to the multilingual BERT model when fine-tuned for a language generation task. Our analysis of TSAuBERT shows that it generally outperforms the most commonly used summarization evaluation metrics in terms of correlation with human assessments. However, the general correlation scores are not significantly large, thus leaving room for further improvements on the summarization evaluation protocol in general.