Automatic extraction of BEL-statements based on neural networks
The automatic extraction of biomedical relations and entities from text has become extremely important in systems biology. For coding the extracted information, the Biological Expression Language (BEL) can be used. A BEL-statement consists of a subject (entity), a predicate (type of relationship), and an object (entity or a further BEL-statement). This paper describes a system based on neural networks (NNs) to extract BEL-statements in the context of the BioCreAtivE 2017 track 3 (task 1) challenge. In our approach, the overall problem is divided into four subtasks: (i) the detection of named entities (NER), (ii) deciding whether a pair of entities participate in a relation, (iii) determining which of the entities participating in a relation is the subject/object entity, and (iv) extracting the type of the relation. By merging the solutions of the subtasks, the BEL-statements are generated. Except for the named entity recognition, (convolutional) NNs were used to solve the tasks. The results show that a neural net based approach is reasonable to use for the extraction of biomedical relations. The limitations of our system are related to the small size (compared to other NN-based applications) of the data set. We argue that by overcoming this limitation, promising results can be expected from NN-based approaches in future.