Desiring text summaries without marketing "hype"

[mbeddo]

Here is the prompt I’m using:

Write an unbiased business summary using the text below. Use American spellings for words like "tumor" and "color". Exclude hype, promotional language, technical language, scientific terminology. Use verbs such as "manufactures", "produces", "distributes", "retails", "sells", "supplies", "leases", "develops", "focuses on", "specializes in", "provides", or "offers". Use nouns such as "products", "services", "solutions", "equipment", or "capabilities". Organize into five parts: Part1 should be a one sentence general overview of the company and the industries/markets it operates within. Part2 should describe the purposes of the company's products, services, solutions, or capabilities. Part2 should begin with "The Company" and use "It" to refer to the company elsewhere. Extract the geographical base location of the company into Part3, report "unknown" if not found. Extract other geographical company locations into Part4 as a comma-delimited list, report "unknown" if there are none. Extract the company URL into Part5, report "unknown" if not found:

Here is some sample text for the prompt:

Text: """ Preston Hollow Capital provides specialized impact financing solutions for projects of significant social and economic importance to local communities in the United States. As a team, we bring a decades-long track record of helping communities achieve their financial, sustainability and community impact goals. We do so through a unique partnership model, rigorous and disciplined credit underwriting and creative investment structuring built around delivering speed, certainty, and flexibility to our borrowers. View source version on businesswire.com: https://www.businesswire.com/news/home/20210715005127/en/ Greg May, Preston Hollow Capital 214.389.0835 gmay@phcllc.com Source: Preston Hollow Capital """

And here is what it returns:

Part1: Preston Hollow Capital is a specialized impact financing company operating in the United States. Part2: The Company provides solutions to help communities achieve their financial, sustainability and community impact goals through a unique partnership model, rigorous credit underwriting and creative investment structuring. Part3: Base location: United States Part4: Other locations: Unknown Part5: URL: https://www.businesswire.com/news/home/20210715005127/en/

I am using the Davinci model.

For the most part, I am happy with the results. However, I need to discourage the use of marketing and promotional words/phrases (want neutral language), so words/phrases like specialized, impact, unique, rigorous, and creative should not be used. I believe this is called “prompt conditioning”?

Might there be a better constructed prompt? If one-shot or few-shot examples would help guide the AI, how would I construct/format that?

I’m trying to avoid fine-tuning, as preparing labelled examples is somewhat expensive for my use case.

Grateful for any advice,

Mike Beddo

[Dent]

If the primary concern is neutral language, you might consider a adding “constitutonal” pipeline, outlined here:

How to set parameters to only answer prompts related to landscaping for example or plumbing etc?

This video describes a pipeline that Anthropic AI uses to moderate their model. Essentially, you get the query from the user, generate a response, then have a moderation model critique said response using a sort of “constitution”, then revise the response according to said constitution, and finally display the response. [First look - RL-CAI/Madison/Claude (fine-tuned 52B) by Anthropic - Announced Dec/2022 (RLAIF v RLHF)] I haven’t implemented it with GPT-3 yet, but it seems like it…

Basically, before returning the final response, submit the first response to an Instruct model to edit the text according to your language preferences, then deliver the revised response

Also, if you have any issues with the more programmatic “parts” of the prompt, well, they should be programmed. I’m honestly surprised it gave you consistent results with a prompt that long. If it were me programming it (and request frequency isn’t super-high), I’d probably submit each “part” as a separate instruction, or at least separate the summary part from the info extraction parts.