Multilingual Summarization for German Texts

Text summarization is the process of condensing a lengthy text into a more manageable size while keeping the key points. This research paper evaluates two models used for abstractive text summarization of German texts, namely MT5 and BERT2BERT. Both models are freely available on Hugging Face, a database for various machine learning models including pre-trained models for the task of summarization. The models chosen for this topic were already trained on a dataset of German news articles. The Bert2Bert model was additionally also trained in other languages such as French, English, Spanish, Russian, and Turkish. Results show that the MT5 model is achieving better results in terms of Rouge-2 scores with higher recall, precision, and F1 values compared to the BERT2BERT model. The pre-processing was a challenge due to errors in recognizing the German special letter “Umlaute”. Another difficulty was detecting white spaces in the sections. To counter these challenges, the scores are generated by developing a task-specific Python code that used an API to the Elsevier website to search for research articles and extract the abstract and content sections automatically.