Vicomtech at LivingNER 2022
Egileak: Naiara Perez Miguel
Data: 20.09.2022
Abstract
This paper describes the participation of the Vicomtech NLP team in the LivingNER 2022 shared task about detecting and normalising mentions of living beings in clinical texts written in Spanish. We participate in each of the 3 LivingNER tasks, combining multiple approaches and strategies. For task 1 (NER) we use a Transformer-based model to perform sequence labelling. For task 2 (Normalisation) we use Semantic Text Search approaches to relate entity mentions to their taxonomy codes. For task 3 we try two different strategies: a trained multi-label classifier and a zero-shot semantic similarity approach. The results for task 1 and task 2 are high, both in our experiments and in the official evaluation results. For task 1 our system obtains an overall of 95.1% of F1-score. For task 2 our system achieves 93.04% F1-score. Task 3 was the most challenging and the one with the least available training data; the scores obtained by all the participating systems have been extremely low. According to the official results, our systems score more than 10 points above the average of the other participating systems for task 1 and 2.
BIB_text
title = {Vicomtech at LivingNER 2022},
keywds = {
Named Entity Recognition, Clinical Text Coding, NCBI Taxonomy, Spanish Clinical Text
}
abstract = {
This paper describes the participation of the Vicomtech NLP team in the LivingNER 2022 shared task about detecting and normalising mentions of living beings in clinical texts written in Spanish. We participate in each of the 3 LivingNER tasks, combining multiple approaches and strategies. For task 1 (NER) we use a Transformer-based model to perform sequence labelling. For task 2 (Normalisation) we use Semantic Text Search approaches to relate entity mentions to their taxonomy codes. For task 3 we try two different strategies: a trained multi-label classifier and a zero-shot semantic similarity approach. The results for task 1 and task 2 are high, both in our experiments and in the official evaluation results. For task 1 our system obtains an overall of 95.1% of F1-score. For task 2 our system achieves 93.04% F1-score. Task 3 was the most challenging and the one with the least available training data; the scores obtained by all the participating systems have been extremely low. According to the official results, our systems score more than 10 points above the average of the other participating systems for task 1 and 2.
}
date = {2022-09-20},
}
