FullMeSH

FullMeSH indexes biomedical full-text articles with Medical Subject Headings (MeSH) to improve automated MeSH assignment by leveraging sectional analysis, attention-based Convolutional Neural Networks (CNNs), and a learning to rank evidence integration framework that combines sparse and deep semantic representations.

Key Features:

• Sectional Analysis: Segments full-text articles into distinct sections with normalized titles to evaluate the contributions of different parts of an article to MeSH indexing.
• Integrated Evidence Framework: Employs a learning to rank framework that integrates evidence from multiple sections by combining sparse and deep semantic representations.
• Attention-based Convolutional Neural Networks (CNNs): Applies attention-based CNNs to each section to improve prediction accuracy, particularly for infrequent MeSH headings.

Scientific Applications:

• Hypothesis generation and knowledge discovery: Produces more accurate MeSH annotations to support literature-based hypothesis generation and knowledge discovery in biomedical research.
• Literature navigation and categorization: Enhances automated categorization and retrieval of biomedical articles by assigning comprehensive MeSH terms across full texts.

Methodology:

Empirically trained on 1.4 million full-text articles from the PubMed Central Open Access subset; uses sectional segmentation, attention-based CNNs per section, and a learning to rank integration of sparse and deep semantic representations; evaluated with a Micro F-measure of 66.76% on a 10,000-article test set, outperforming DeepMeSH and MeSHLabeler by 3.3% and 6.4% respectively and improving Check Tag indexing versus DeepMeSH by 4.7%.