TLDR: Extreme Summarization of Scientific Documents

Code

• github.com/allenai/scitldr
  OfficialIn papernone★ 759
• github.com/iamgroot42/mimir
  pytorch★ 185
• github.com/jungokasai/twist_decoding
  pytorch★ 30
• github.com/bnitsan/papertweet
  pytorch★ 5

Abstract

We introduce TLDR generation, a new form of extreme summarization, for scientific papers. TLDR generation involves high source compression and requires expert background knowledge and understanding of complex domain-specific language. To facilitate study on this task, we introduce SciTLDR, a new multi-target dataset of 5.4K TLDRs over 3.2K papers. SciTLDR contains both author-written and expert-derived TLDRs, where the latter are collected using a novel annotation protocol that produces high-quality summaries while minimizing annotation burden. We propose CATTS, a simple yet effective learning strategy for generating TLDRs that exploits titles as an auxiliary training signal. CATTS improves upon strong baselines under both automated metrics and human evaluations. Data and code are publicly available at https://github.com/allenai/scitldr.

Tasks

Reproductions