Here is a summary of the key points from the article:
The paper develops a method called Chain of Density (CoD) prompting to generate increasingly dense summaries from GPT-4. The prompt iteratively identifies missing entities from the source text and fuses them into the previous summary without increasing length. This forces the model to become more abstractive and exhibit more fusion.
The authors evaluate CoD summaries on 100 CNN/Dailymail articles. Statistics show the summaries get progressively more dense, abstractive, and incorporate more content from later in the article.
Human evaluations indicate moderate preference for CoD summaries slightly denser than vanilla GPT-4 ones, but not as dense as human references. GPT-4 assessment also favors increased density up to a point before coherence declines.
Qualitative analysis reveals a tradeoff between informativeness (favors more entities) and coherence/readability (favors fewer entities). The middle CoD steps best balance this.
The authors release 500 annotated and 5,000 unannotated CoD summaries to further study controllable density in summarization.
