Memory-Efficient MeSH Tagging with Character-Based Hierarchical Context Classification
▻https://research.science.ai/article/memory-efficient-mesh-tagging
Medical Subject Headings (MeSH) represents a monumental effort in categorizing the breadth and depth of concepts within the biomedical sciences, and has been instrumental in improving the indexing of biomedical scholarly articles. There have been many efforts over the years to create a completely automated MeSH tagging system, but such a system has proven to be quite challenging. This is in no small part due to the vast scope of MeSH. Many of these systems require large resource footprints due to the necessity of performing computation on large sets of documents, or over large numbers of iterations or steps. We present here a memory-efficient automatic MeSH tagging system written entirely in JavaScript that is lightweight enough to be run in the web browser, using hierarchical context classification with character-based convolutional and recurrent neural networks. These deep neural networks can be run in modern web browsers or in Node.js through neocortex.js, an open source library developed as a part of this work. The system is evaluated on over 15 million abstracts in MEDLINE/PubMED, with resulting performance similar to that of existing systems, but at a significantly reduced computational cost and model complexity.
