Improved Biomedical Word Embeddings in the Transformer Era

Biomedical word embeddings are usually pre-trained on free text corpora with neural methods that capture local and global distributional properties. They are leveraged in downstream tasks using various neural architectures that are designed to optimize task-specific objectives that might further tune such embeddings. Since 2018, however, there is a marked shift from these static embeddings to contextual embeddings motivated by language models (e.g., ELMo, transformers such as BERT, and ULMFiT). These dynamic embeddings have the added benefit of being able to distinguish homonyms and acronyms given their context. However, static embeddings are still relevant in low resource settings (e.g., smart devices, IoT elements) and to study lexical semantics from a computational linguistics perspective. In this paper, we jointly learn word and concept embeddings by first using the skip-gram method and further fine-tuning them with correlational information manifesting in co-occurring Medical Subject Heading (MeSH) concepts in biomedical citations. This fine-tuning is accomplished with the BERT transformer architecture in the two-sentence input mode with a classification objective that captures MeSH pair co-occurrence. In essence, we repurpose a transformer architecture (typically used to generate dynamic embeddings) to improve static embeddings using concept correlations. We conduct evaluations of these tuned static embeddings using multiple datasets for word relatedness developed by previous efforts. Without selectively culling concepts and terms (as was pursued by previous efforts), we believe we offer the most exhaustive evaluation of static embeddings to date with clear performance improvements across the board. We provide our code and embeddings for public use for downstream applications and research endeavors: https://github.com/bionlproc/BERT-CRel-Embeddings