From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting. Adams, G., Fabbri, A., Ladhak, F., Lehman, E., & Elhadad, N. September, 2023. arXiv:2309.04269 [cs]![From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting [link]](https://bibbase.org/img/filetypes/link.svg "link")
Paper doi abstract bibtex Selecting the ``right'' amount of information to include in a summary is a difficult task. A good summary should be detailed and entity-centric without being overly dense and hard to follow. To better understand this tradeoff, we solicit increasingly dense GPT-4 summaries with what we refer to as a ``Chain of Density'' (CoD) prompt. Specifically, GPT-4 generates an initial entity-sparse summary before iteratively incorporating missing salient entities without increasing the length. Summaries generated by CoD are more abstractive, exhibit more fusion, and have less of a lead bias than GPT-4 summaries generated by a vanilla prompt. We conduct a human preference study on 100 CNN DailyMail articles and find that that humans prefer GPT-4 summaries that are more dense than those generated by a vanilla prompt and almost as dense as human written summaries. Qualitative analysis supports the notion that there exists a tradeoff between informativeness and readability. 500 annotated CoD summaries, as well as an extra 5,000 unannotated summaries, are freely available on HuggingFace (https://huggingface.co/datasets/griffin/chain_of_density).
@misc{adams_sparse_2023,
title = {From {Sparse} to {Dense}: {GPT}-4 {Summarization} with {Chain} of {Density} {Prompting}},
shorttitle = {From {Sparse} to {Dense}},
url = {http://arxiv.org/abs/2309.04269},
doi = {10.48550/arXiv.2309.04269},
abstract = {Selecting the ``right'' amount of information to include in a summary is a difficult task. A good summary should be detailed and entity-centric without being overly dense and hard to follow. To better understand this tradeoff, we solicit increasingly dense GPT-4 summaries with what we refer to as a ``Chain of Density'' (CoD) prompt. Specifically, GPT-4 generates an initial entity-sparse summary before iteratively incorporating missing salient entities without increasing the length. Summaries generated by CoD are more abstractive, exhibit more fusion, and have less of a lead bias than GPT-4 summaries generated by a vanilla prompt. We conduct a human preference study on 100 CNN DailyMail articles and find that that humans prefer GPT-4 summaries that are more dense than those generated by a vanilla prompt and almost as dense as human written summaries. Qualitative analysis supports the notion that there exists a tradeoff between informativeness and readability. 500 annotated CoD summaries, as well as an extra 5,000 unannotated summaries, are freely available on HuggingFace (https://huggingface.co/datasets/griffin/chain\_of\_density).},
urldate = {2024-02-13},
publisher = {arXiv},
author = {Adams, Griffin and Fabbri, Alexander and Ladhak, Faisal and Lehman, Eric and Elhadad, Noémie},
month = sep,
year = {2023},
note = {arXiv:2309.04269 [cs]},
keywords = {/unread, Computer Science - Computation and Language},
}
Downloads: 0
{"_id":"5wzQHFfnRhBTC5ACm","bibbaseid":"adams-fabbri-ladhak-lehman-elhadad-fromsparsetodensegpt4summarizationwithchainofdensityprompting-2023","author_short":["Adams, G.","Fabbri, A.","Ladhak, F.","Lehman, E.","Elhadad, N."],"bibdata":{"bibtype":"misc","type":"misc","title":"From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting","shorttitle":"From Sparse to Dense","url":"http://arxiv.org/abs/2309.04269","doi":"10.48550/arXiv.2309.04269","abstract":"Selecting the ``right'' amount of information to include in a summary is a difficult task. A good summary should be detailed and entity-centric without being overly dense and hard to follow. To better understand this tradeoff, we solicit increasingly dense GPT-4 summaries with what we refer to as a ``Chain of Density'' (CoD) prompt. Specifically, GPT-4 generates an initial entity-sparse summary before iteratively incorporating missing salient entities without increasing the length. Summaries generated by CoD are more abstractive, exhibit more fusion, and have less of a lead bias than GPT-4 summaries generated by a vanilla prompt. We conduct a human preference study on 100 CNN DailyMail articles and find that that humans prefer GPT-4 summaries that are more dense than those generated by a vanilla prompt and almost as dense as human written summaries. Qualitative analysis supports the notion that there exists a tradeoff between informativeness and readability. 500 annotated CoD summaries, as well as an extra 5,000 unannotated summaries, are freely available on HuggingFace (https://huggingface.co/datasets/griffin/chain_of_density).","urldate":"2024-02-13","publisher":"arXiv","author":[{"propositions":[],"lastnames":["Adams"],"firstnames":["Griffin"],"suffixes":[]},{"propositions":[],"lastnames":["Fabbri"],"firstnames":["Alexander"],"suffixes":[]},{"propositions":[],"lastnames":["Ladhak"],"firstnames":["Faisal"],"suffixes":[]},{"propositions":[],"lastnames":["Lehman"],"firstnames":["Eric"],"suffixes":[]},{"propositions":[],"lastnames":["Elhadad"],"firstnames":["Noémie"],"suffixes":[]}],"month":"September","year":"2023","note":"arXiv:2309.04269 [cs]","keywords":"/unread, Computer Science - Computation and Language","bibtex":"@misc{adams_sparse_2023,\n\ttitle = {From {Sparse} to {Dense}: {GPT}-4 {Summarization} with {Chain} of {Density} {Prompting}},\n\tshorttitle = {From {Sparse} to {Dense}},\n\turl = {http://arxiv.org/abs/2309.04269},\n\tdoi = {10.48550/arXiv.2309.04269},\n\tabstract = {Selecting the ``right'' amount of information to include in a summary is a difficult task. A good summary should be detailed and entity-centric without being overly dense and hard to follow. To better understand this tradeoff, we solicit increasingly dense GPT-4 summaries with what we refer to as a ``Chain of Density'' (CoD) prompt. Specifically, GPT-4 generates an initial entity-sparse summary before iteratively incorporating missing salient entities without increasing the length. Summaries generated by CoD are more abstractive, exhibit more fusion, and have less of a lead bias than GPT-4 summaries generated by a vanilla prompt. We conduct a human preference study on 100 CNN DailyMail articles and find that that humans prefer GPT-4 summaries that are more dense than those generated by a vanilla prompt and almost as dense as human written summaries. Qualitative analysis supports the notion that there exists a tradeoff between informativeness and readability. 500 annotated CoD summaries, as well as an extra 5,000 unannotated summaries, are freely available on HuggingFace (https://huggingface.co/datasets/griffin/chain\\_of\\_density).},\n\turldate = {2024-02-13},\n\tpublisher = {arXiv},\n\tauthor = {Adams, Griffin and Fabbri, Alexander and Ladhak, Faisal and Lehman, Eric and Elhadad, Noémie},\n\tmonth = sep,\n\tyear = {2023},\n\tnote = {arXiv:2309.04269 [cs]},\n\tkeywords = {/unread, Computer Science - Computation and Language},\n}\n\n","author_short":["Adams, G.","Fabbri, A.","Ladhak, F.","Lehman, E.","Elhadad, N."],"key":"adams_sparse_2023-1","id":"adams_sparse_2023-1","bibbaseid":"adams-fabbri-ladhak-lehman-elhadad-fromsparsetodensegpt4summarizationwithchainofdensityprompting-2023","role":"author","urls":{"Paper":"http://arxiv.org/abs/2309.04269"},"keyword":["/unread","Computer Science - Computation and Language"],"metadata":{"authorlinks":{}},"downloads":0,"html":""},"bibtype":"misc","biburl":"https://bibbase.org/zotero/alexphoto","dataSources":["xFKgoYWMc8jcB9rcq"],"keywords":["/unread","computer science - computation and language"],"search_terms":["sparse","dense","gpt","summarization","chain","density","prompting","adams","fabbri","ladhak","lehman","elhadad"],"title":"From Sparse to Dense: GPT-4 Summarization with Chain of Density Prompting","year":2023}