You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Selecting the right'' amount of information to include in a summary is adifficult task. A good summary should be detailed and entity-centric withoutbeing overly dense and hard to follow. To better understand this tradeoff, wesolicit increasingly dense GPT-4 summaries with what we refer to as a Chainof Density'' (CoD) prompt. Specifically, GPT-4 generates an initialentity-sparse summary before iteratively incorporating missing salient entitieswithout increasing the length. Summaries generated by CoD are more abstractive,exhibit more fusion, and have less of a lead bias than GPT-4 summariesgenerated by a vanilla prompt. We conduct a human preference study on 100 CNNDailyMail articles and find that that humans prefer GPT-4 summaries that aremore dense than those generated by a vanilla prompt and almost as dense ashuman written summaries. Qualitative analysis supports the notion that thereexists a tradeoff between informativeness and readability. 500 annotated CoDsummaries, as well as an extra 5,000 unannotated summaries, are freelyavailable on HuggingFace(https://huggingface.co/datasets/griffin/chain_of_density).
AkihikoWatanabe
changed the title
あ
From Sparse to Dense: GPT-4 Summarization with Chain of Density
Prompting, Griffin Adams+, N/A, arXiv'23
Sep 17, 2023
URL
Affiliations
Abstract
right'' amount of information to include in a summary is adifficult task. A good summary should be detailed and entity-centric withoutbeing overly dense and hard to follow. To better understand this tradeoff, wesolicit increasingly dense GPT-4 summaries with what we refer to as a
Chainof Density'' (CoD) prompt. Specifically, GPT-4 generates an initialentity-sparse summary before iteratively incorporating missing salient entitieswithout increasing the length. Summaries generated by CoD are more abstractive,exhibit more fusion, and have less of a lead bias than GPT-4 summariesgenerated by a vanilla prompt. We conduct a human preference study on 100 CNNDailyMail articles and find that that humans prefer GPT-4 summaries that aremore dense than those generated by a vanilla prompt and almost as dense ashuman written summaries. Qualitative analysis supports the notion that thereexists a tradeoff between informativeness and readability. 500 annotated CoDsummaries, as well as an extra 5,000 unannotated summaries, are freelyavailable on HuggingFace(https://huggingface.co/datasets/griffin/chain_of_density).Translation (by gpt-3.5-turbo)
Summary (by gpt-3.5-turbo)
The text was updated successfully, but these errors were encountered: