Vicomtech at MESINESP2: BERT-based Multi-label Classification Models for Biomedical Text Indexing
Autores: Naiara Perez Miguel
Fecha: 22.09.2021
Abstract
This paper describes the participation of the Vicomtech NLP team in the MESINESP2 shared task. The challenge consists in the development of systems for the automatic indexing with DeCS codes of healthrelated documents in Spanish. The systems submitted by Vicomtech are multilabel classifiers based on pre-trained BERT models. We have experimented with multiple ways of representing the documents, such as encoding DeCS term glosses along with the input text. According to the official evaluation
results, our systems are surpassed by other competing teams—despite being fast and achieving good precision, we fall behind especially in recall metrics. Overall, the task remains challenging even for the best performing systems and there is ample room to advance the state of the art for this particular task.
BIB_text
title = {Vicomtech at MESINESP2: BERT-based Multi-label Classification Models for Biomedical Text Indexing},
pages = {221-230},
keywds = {
Biomedical, BERT, Classification
}
abstract = {
This paper describes the participation of the Vicomtech NLP team in the MESINESP2 shared task. The challenge consists in the development of systems for the automatic indexing with DeCS codes of healthrelated documents in Spanish. The systems submitted by Vicomtech are multilabel classifiers based on pre-trained BERT models. We have experimented with multiple ways of representing the documents, such as encoding DeCS term glosses along with the input text. According to the official evaluation
results, our systems are surpassed by other competing teams—despite being fast and achieving good precision, we fall behind especially in recall metrics. Overall, the task remains challenging even for the best performing systems and there is ample room to advance the state of the art for this particular task.
}
date = {2021-09-22},
}
