I am researching potential ways to create better train-
ing data and training strategies for transformer-based deep learning models, with primary emphasis
on the NER task and large-scale application to English biomedical scientific text.

The research topics include data augmentation for NER to increase the classification per-
formance of NER models, combining already available but separate NER resources for
training models that simultaneously predict multiple named entity types (as biomedical
resources often consider only one entity type at a time) and utilizing large-scale data in
setups with distant or weak supervision from noisy and automatically classified data to
create new or enhance the available data sets and increase the performance of classifiers.
The research will be carried out as part of a four-year project funded by the Academy
of Finland that is in collaboration with University of Copenhagen researchers who de-
velop the largest biomedical text mining resource, STRING database. The primary
practical outcome of the work is to be able to efficiently (both in tagging performance
and run-time performance) use the methods planned in this research to improve this key
biomedical resource.