Reproduce “AstroConcepts: A Large-Scale Multi-Label Classification Corpus for Astrophysics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Atilla Kaan Alkan, Felix Grezes, Sergi Blanco-Cuaresma, Jennifer Lynn Bartlett, Daniel Chivvis, Anna Kelbert, Kelly Lockhart, Alberto Accomazzi
Researchers created AstroConcepts, a large corpus for multi-label classification in astrophysics, featuring over 21,000 abstracts annotated with 1,800+ hierarchically structured concepts. A hybrid approach, combining a domain-adapted supervised model with a vocabulary-constrained Large Language Model, demonstrated superior performance, achieving a Macro-F1 of 0.377 and notable improvements for rare concepts.
Key insights
- Domain adaptation yields asymmetric benefits, offering greater *relative* performance improvements for rare ("Tail") concepts, suggesting targeted value for specialized terminology.
- Classification performance consistently peaks for mid-frequency ("Torso") concepts across all methods, highlighting inherent challenges at the extremes of the label frequency distribution.
- Vocabulary-constrained LLMs demonstrate a parameter-efficient pathway, achieving strong overall performance and substantially better classification of rare concepts by integrating structured domain knowledge.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.