Feature Importance for Biomedical Named Entity Recognition

Abstract

Within the domain of biomedical natural language processing (bioNLP), researchers have used many token features for machine learning models. With recent progress in word embeddings algorithms, it is no longer clear if most of these features are still useful. In this paper we survey the features which have been used in bioNLP, and evaluate each feature’s utility in a sample bioNLP task: the N2C2 2018 named entity recognition challenge. The features we test include two types of word embeddings, syntactic, lexical, and orthographic features, character-embeddings, and clustering and distributional word representations. We find that using fastText word embeddings results in a significantly higher F1 score than using any other individual feature (0.9142 compared to 0.8750 for the next-best feature). Furthermore, we conducted several experiments using combinations of features, and found that all tested combinations attained a lower F1 score than using word embeddings only. This indicates that supplementing word embeddings with additional features is not beneficial, and may even be detrimental.