As a result of technological progress (hardware and software), the amount of data is increasing and we are working to include all information that could be of importance. The task from the field of NLP that tries to automate this process, information extraction (IE), or rather, its subtask, relation extraction (RE), is studied in combination with the state of the art "Bidirectional Encoder Representations from Transformers" (BERT). In this thesis, first the basic concepts of Deep Learning and Natural Language Processing (NLP) with a brief historical overview are introduced. Then, the summary on datasets suitable for the task of extracting sentence-level relations is presented. Based on the presented datasets, a new dataset for scientific RE, combo160, is created and used to fine-tune the BERT and SciBERT models for the task. From the results, it can be inferred that thematically similar (here: scientific) pretrainig corpora can indeed improve the performance of the later fine- tuned models for RE on scientific data.