KU-DMIS at BioASQ 9: Data-centric and model-centric approaches for biomedical question answering

In this paper, we present approaches for our participation in the 9th BioASQ challenge (Task b - Phase B). Our systems are based on the transformer models with model-centric and data-centric approaches. For factoid-type questions we modified the dataset to increase label consistency, and for list-type questions we apply the sequence tagging model which is a more natural model design for the multi-label task. Our experimental results suggest two main points: better model design can be achieved by reflecting data characteristics such as the number of labels for a data point; and scarce resources such as BioQA datasets can greatly benefit from a data-centric approach with relatively little effort. Our submissions achieve competitive results with top or near top performance in the challenge.