Study: Evaluation design impacts MeSH feature performance gap

A new study published on arXiv investigates the impact of evaluation design on the performance gap between expert-assigned and automatically generated Medical Subject Headings (MeSH) when used as features in classification tasks. The research compared a bag-of-words logistic regression model with BiomedBERT, a domain-specific language model, on the Cohen benchmark for drug-class identification. Findings indicate that the observed gap between expert and auto-assigned MeSH can vary significantly based on evaluation methodologies, such as corpus size and cross-validation folds. The study also noted that transformer models like BiomedBERT may face token limits with appended MeSH terms, potentially affecting their performance. AI

RANK_REASON Academic paper published on arXiv detailing experimental results and analysis. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.CL →

• arXiv
• attention deficit hyperactivity disorder
• Cohen
• Domain-Specific Language Model Pretraining for Biomedical Natural Language Processing
• Hugging Face
• Medical Subject Headings
• Opioids
• statin

• paper
• other

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →