Humboldt @ DrugProt: Chemical-Protein Relation Extraction with Pretrained Transformers and Entity Descriptions

The detection of chemical-protein interactions is an important task with applications in drug design and biotechnology. The BioCreative VII DrugProt shared task provides a benchmark for the automated extraction of such relations from scientific text. This article describes the Humboldt approach to solving it. We define the task as a relation classification problem, which we model with pretrained transformer language models and further use entity descriptions as an additional knowledge source. On the hidden test set of DrugProt, our model achieves 79.73% F1, yielding an improvement of over 17pp over the average score of all task participants. Keywords—relation extraction; transformers; entity descriptions