Polimi at CLinkaRT: a Conditional Random Field vs a BERT-based approach

In the context of the EVALITA 2023 challenge, we present the models we have developed for the CLinkaRT task, which aims to identify medical examinations and their corresponding results in Italian clinical documents. We propose two distinct approaches: one utilising a Conditional Random Field (CRF), a probabilistic graphical model traditionally used for Named Entity Recognition, and the other based on BERT, the transformer-based model that is currently state-of-the-art for many Natural Language Processing tasks. Both models incorporate external knowledge from publicly available medical resources and are enhanced with heuristic rules to establish associations between exams and results. Our comparative analysis elects the CRF-based model as the winner, securing the third position in the competition ranking, but the BERT-based model demonstrated competitive performance.