Biomedical Event Extraction by Multi-class Classification of Pairs of Text Entities
Abstract
This paper describes the HDS4NLP en- try to the BioNLP 2013 shared task on biomedical event extraction. This system is based on a pairwise model that transforms trigger classification in a simple multi-class problem in place of the usual multi-label problem. This model facilitates inference compared to global models while relying on richer information compared to usual pipeline approaches. The HDS4NLP system ranked 6th on the Genia task (43.03% f-score), and after fixing a bug discovered after the final submission, it outperforms the winner of this task (with a f-score of 51.15%).
Origin : Publisher files allowed on an open archive
Loading...