Data-Centric and Model-Centric Approaches for Biomedical Question Answering

Biomedical question answering (BioQA) is the process of automated information extraction from the biomedical literature, and as the number of accessible biomedical papers is increasing rapidly, BioQA is attracting more attention. In order to improve the performance of BioQA systems, we designed strategies for the sub-tasks of BioQA and assessed their effectiveness using the BioASQ dataset. We designed data-centric and model-centric strategies based on the potential for improvement for each sub-task. For example, model design for the factoid-type questions has been explored intensely but the potential of increased label consistency has not been investigated (data-centric approach). On the other hand, for list-type questions, we apply the sequence tagging model as it is more natural for the multi-answer (i.e. multi-label) task (model-centric approach). Our experimental results suggest two main points: scarce resources like BioQA datasets can be benefited from data-centric approaches with relatively little effort; and a model design reflecting data characteristics can improve the performance of the system. The scope of this paper is majorly focused on applications of our strategies in the BioASQ 8b dataset and our participating systems in the 9th BioASQ challenges. Our submissions achieve competitive results with top or near top performance in the 9th challenge (Task b - Phase B).