Skip to content

Repository files navigation

Abstract

High-quality scientific extreme summary (TLDR) facilitates effective science communication. How do large language models (LLMs) perform in generating them? How are LLM-generated summaries different from those written by human experts? However, the lack of a comprehensive, high-quality scientific TLDR dataset hinders both the development and evaluation of LLMs' summarization ability. To address these, we propose a novel dataset, BiomedTLDR, containing a large sample of researcher-authored summaries from scientific papers, which leverages the common practice of including authors' comments alongside bibliography items. We then test popular open-weight LLMs for generating TLDRs based on abstracts. Our analysis reveals that, although some of them successfully produce humanoid summaries, LLMs generally exhibit a greater affinity for the original text's lexical choices and rhetorical structures, hence tend to be more extractive rather than abstractive in general, compared to humans.

File location

Data

  1. LLM-generated summaries: data/paper_html_10.1038/abs_annotation/generated_annotations

  2. Human-authored summaries: data/BiomedTLDR_dataset.tsv or check our dataset at Hugging Face

Experiment code

  1. LLMs are more extractive summarizer: evaluation/ref_based
  2. LLMs outperform human consistently: evaluation/predict_task
  3. LLMs mirror human’s citation bias: bias

About

Repository for the paper "LLMs exhibit better capability in generating extreme summarization"

Resources

Stars

Watchers

Forks

Contributors

Languages