Medical Document Summarization
Comparing T5, BART, and PEGASUS for abstractive summarization of medical research documents.
Problem. Many medical research documents don’t come with a summary or abstract, making it hard for researchers to quickly judge relevance without reading the whole document.
Approach. We compare five abstractive summarization models — T5 (small, base, large), BART, and PEGASUS — on the SUMPUBMED dataset (32,689 PubMed documents), scored with ROUGE-1/2/L.
Results. PEGASUS comes out on top on most metrics (ROUGE-1 recall 0.37, F1 0.36), with BART close behind. Published at ICAAIC 2023 (Evani et al., 2023).
Links. Project page · Paper
ROUGE-1/2/L recall, precision, and F1 across T5 (small/base/large), BART, and PEGASUS.