BioMegatron: Larger Biomedical Domain Language Model
This work addresses the need for better biomedical NLP tools for researchers and practitioners, but it is incremental as it builds on existing domain-specific models by scaling up size and data.
The paper tackles the problem of improving biomedical domain language models by empirically studying factors like model size and pre-training corpus, showing that a larger model (BioMegatron) trained on more data achieves noticeable improvements over previous state-of-the-art on benchmarks such as named entity recognition, relation extraction, and question answering.
There has been an influx of biomedical domain-specific language models, showing language models pre-trained on biomedical text perform better on biomedical domain benchmarks than those trained on general domain text corpora such as Wikipedia and Books. Yet, most works do not study the factors affecting each domain language application deeply. Additionally, the study of model size on domain-specific models has been mostly missing. We empirically study and evaluate several factors that can affect performance on domain language applications, such as the sub-word vocabulary set, model size, pre-training corpus, and domain transfer. We show consistent improvements on benchmarks with our larger BioMegatron model trained on a larger domain corpus, contributing to our understanding of domain language model applications. We demonstrate noticeable improvements over the previous state-of-the-art (SOTA) on standard biomedical NLP benchmarks of named entity recognition, relation extraction, and question answering. Model checkpoints and code are available at [https://ngc.nvidia.com] and [https://github.com/NVIDIA/NeMo].