Harnessing the Power of AI to Rapidly Extract Key Insights using CoD | by Raphael Mansuy | Sep, 2023

Learning and digesting prolonged articles, evaluation papers, or books could be time-consuming.

Recently, I bought right here all through an intriguing new methodology known as Chain of Density (CoD) that leverages huge language fashions like GPT-3 to provide dense, informative summaries.

On this put up, I’ll make clear how CoD works and share code for producing summaries with this method.

Making a summary that completely captures the necessary factor particulars of a provide doc, whereas remaining easy to be taught and digest, is a hard balancing act. Summaries that are too sparse miss very important knowledge. Nonetheless pack an extreme quantity of ingredient proper right into a small home, and the summary turns into dense and sophisticated.

One of the best summary hits a sweet spot — dense adequate to be informative, however gentle adequate to be taught simply. This steadiness is especially very important when dealing with prolonged articles or papers. You want to boil down the necessary factor particulars into a good summary you’ll have the ability to evaluation shortly.

Recently, researchers at Columbia Faculty, MIT, and Salesforce revealed a paper describing Chain of Density (CoD), a model new methodology for producing optimally dense GPT-based summaries. CoD takes advantage of GPT-3’s few-shot prompting capabilities to iteratively densify an preliminary sparse summary.

Proper right here’s an abstract of how CoD works:

1. GPT-4 generates an preliminary verbose, entity-sparse summary. This summary intentionally consists of minimal specific particulars — merely adequate context to make sense.
2. GPT-4 then identifies 1–3 “Missing Entities” from the provision doc that are associated, specific, novel (not already coated), reliable to the distinctive content material materials, and concise (beneath 5 phrases).
3. Subsequent, GPT-4 rewrites the sooner summary to fuse throughout the Missing Entities contained in the similar phrase limit. This forces abstraction, fusion, and compression to make room for the additional particulars.
4. Steps 2–3 are repeated as a lot as 5 events, resulting in an increasing number of dense summaries with further entities coated per token.