Effects of Semantic Features on Machine Learning-Based Drug Name Recognition Systems: Word Embeddings vs. Manually Constructed Dictionaries

Abstract

Semantic features are very important for machine learning-based drug name recognition (DNR) systems. The semantic features used in most DNR systems are based on drug dictionaries manually constructed by experts. Building large-scale drug dictionaries is a time-consuming task and adding new drugs to existing drug dictionaries immediately after they are developed is also a challenge. In recent years, word embeddings that contain rich latent semantic information of words have been widely used to improve the performance of various natural language processing tasks. However, they have not been used in DNR systems. Compared to the semantic features based on drug dictionaries, the advantage of word embeddings lies in that learning them is unsupervised. In this paper, we investigate the effect of semantic features based on word embeddings on DNR and compare them with semantic features based on three drug dictionaries. We propose a conditional random fields (CRF)-based system for DNR. The skip-gram model, an unsupervised algorithm, is used to induce word embeddings on about 17.3 GigaByte (GB) unlabeled biomedical texts collected from MEDLINE (National Library of Medicine, Bethesda, MD, USA). The system is evaluated on the drug-drug interaction extraction (DDIExtraction) 2013 corpus. Experimental results show that word embeddings significantly improve the performance of the DNR system and they are competitive with semantic features based on drug dictionaries. F-score is improved by 2.92 percentage points when word embeddings are added into the baseline system. It is comparative with the improvements from semantic features based on drug dictionaries. Furthermore, word embeddings are complementary to the semantic features based on drug dictionaries. When both word embeddings and semantic features based on drug dictionaries are added, the system achieves the best performance with an F-score of 78.37%, which outperforms the best system of the DDIExtraction 2013 challenge by 6.87 percentage points.