Reproduce “Generating Summaries with Controllable Readability Levels” 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
Leonardo F. R. Ribeiro, Mohit Bansal, Markus Dreyer
This research introduces methods for generating abstractive summaries with fine-grained, continuous control over readability levels. The lookahead decoding approach achieves the most precise control, reducing the average absolute difference from target Flesch Reading Ease (FRE∆) to 4.5 and boosting the correlation with the target to 0.95, outperforming other methods and a GPT3.5 baseline.
Key insights
- Fine-grained, continuous control over readability in abstractive summarization is effectively achieved through instruction-based training, reinforcement learning, and especially a lookahead decoding strategy.
- There is an inherent trade-off between readability and other crucial summary attributes; easier-to-read summaries tend to be less specific, less factually consistent, and less coherent.
- Human evaluations confirm the perceived readability differences, while also suggesting that increased readability can sometimes lead to a reduction in perceived informativeness.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.