GPT-3 Prompts for Sentence-Level and Paragraph-Level Text Summarization / Text Shortening / Text Rewriting

Question

Need effective prompts for GPT-3 that can accomplish this 'programming' task. Creating effective GPT-3 prompts has essentially become a new form of programming (giving a computer instructions to complete a task).

There are getting to be repositories for the nascient, growing 'programming' language of GPT-3 prompts, eg at:

https://github.com/martonlanga/gpt3-prompts

http://gptprompts.wikidot.com/start

https://github.com/wgryc/gpt3-prompts

See a working example below, which works ok, but doesn't really address the need, and isn't adequately reliable.

This is an important, new, and quickly growing area.

Seeking prompts that will accomplish the goal in the Title: summarizing / shortening sentences and / or paragraphs with high reliability, without creating nonsense.

Please, reviewers, this is an important question to many people... don't be narrow-minded and decided that because GPT-3 prompts aren't (yet) a 'traditional' computer language they don't have a place here.

Thank you for your help

Example GPT-3 Prompt:

Please summarize the article below. """ Microsoft in talks to buy TikTok Negotiations for ByteDance-owned social media group come as Trump threatens action

Microsoft has held talks to acquire TikTok, whose Chinese owner ByteDance faces mounting pressure from the US government to sell the video sharing app or risk being blacklisted in the country, said people briefed on the matter.

... the rest of the article... """

Q: Could you please summarize the article above in three sentences?

Answer 1

Q: Could you please summarize the article above in three sentences?

GPT-3 does offer this functionality from the get-go. The Davici engine is best suited for summarizations (though to the cost of speed - Curie and Cushman are safer, but more inaccurate on the other hand).

What follows is an example. I took a snippet of the Wikipedia article about Artificial Intelligence and didn't remove formatting whatsoever. Using davinci-instruct-beta and a maximum response length of 100 tokens for a "short summary":

Input:

Artificial intelligence (AI) is intelligence demonstrated by machines, unlike the natural intelligence displayed by humans and animals, which involves consciousness and emotionality. The distinction between the former and the latter categories is often revealed by the acronym chosen. 'Strong' AI is usually labelled as AGI (Artificial General Intelligence) while attempts to emulate 'natural' intelligence have been called ABI (Artificial Biological Intelligence). Leading AI textbooks define the field as the study of "intelligent agents": any device that perceives its environment and takes actions that maximize its chance of successfully achieving its goals.3 Colloquially, the term "artificial intelligence" is often used to describe machines (or computers) that mimic "cognitive" functions that humans associate with the human mind, such as "learning" and "problem solving".[4]

As machines become increasingly capable, tasks considered to require "intelligence" are often removed from the definition of AI, a phenomenon known as the AI effect.[5] A quip in Tesler's Theorem says "AI is whatever hasn't been done yet."[6] For instance, optical character recognition is frequently excluded from things considered to be AI,[7] having become a routine technology.[8] Modern machine capabilities generally classified as AI include successfully understanding human speech,[9] competing at the highest level in strategic game systems (such as chess and Go),[10] autonomously operating cars, intelligent routing in content delivery networks, and military simulations.[11]

tl;dr

Output:

AI is intelligence demonstrated by machines, unlike the natural intelligence displayed by humans and animals, which involves consciousness and emotionality.

You could use Python (and many other programming languages) to call this API from your workflow, in this case with the following code snippet:

import openai

openai.api_key 'KEY'

response = openai.Completion.create(
  engine="davinci-instruct-beta",
  prompt="Artificial intelligence (AI) is intelligence demonstrated by machines, unlike the natural intelligence displayed by humans and animals, which involves consciousness and emotionality. The distinction between the former and the latter categories is often revealed by the acronym chosen. 'Strong' AI is usually labelled as AGI (Artificial General Intelligence) while attempts to emulate 'natural' intelligence have been called ABI (Artificial Biological Intelligence). Leading AI textbooks define the field as the study of \"intelligent agents\": any device that perceives its environment and takes actions that maximize its chance of successfully achieving its goals.[3] Colloquially, the term \"artificial intelligence\" is often used to describe machines (or computers) that mimic \"cognitive\" functions that humans associate with the human mind, such as \"learning\" and \"problem solving\".[4]\n\nAs machines become increasingly capable, tasks considered to require \"intelligence\" are often removed from the definition of AI, a phenomenon known as the AI effect.[5] A quip in Tesler's Theorem says \"AI is whatever hasn't been done yet.\"[6] For instance, optical character recognition is frequently excluded from things considered to be AI,[7] having become a routine technology.[8] Modern machine capabilities generally classified as AI include successfully understanding human speech,[9] competing at the highest level in strategic game systems (such as chess and Go),[10] autonomously operating cars, intelligent routing in content delivery networks, and military simulations.[11]\n\ntl;dr:",
  temperature=0.25,
  max_tokens=100,
  top_p=1
)

response now holds the above mentioned output.

---

In your specific case you wanted a summarization in three sentences, so here you go:

From my subjective point of view, that is a very precise explanation of the article!

Answer 2

Unfortunately there is not a way in GPT-3 that the instruction will be upheld within our requirement. Remember that GPT-3 is non-deterministic. Even the response length parameter is not strictly followed (except a possible truncation at maximum length). Instead of procedural, deterministic result, the GPT-3 value to me is to navigate the vast human language and produce sensible result. I say it looks like an actor: looks good, talks good, but don't trust it 100%.

What I suggest you to do is:

1. experiment with the prompt that is having most likely result, then
2. guard the output using traditional programming. e.g. if you want no more than 3 sentences (strictly), then you need to parse/split the output and discard anything that is not complying to your restrictions (post-process). Then you may need to issue another request.

The preprocessing or post-processing of the prompt/completion is a way to "tame" GPT-3 other than setting the temperature and/or top P. Of course there are endless pre/post-processing algorithm/model/NLP methods that can be used here.

Answer 3

In my experience prompts that form a single text with the expected output often work better than prompts consisting of context + question/order. So instead of writing

[Some text]

Could you please summarize the article above in three sentences?

I would try

[Some text]

Summary in three sentences:

I would also suggest dropping all the non-pertinent words (Could you, please, etc) from the question in prompt.

Another useful technique is to provide some examples within the prompt before the question to better tune the output of the net:

This is the text:
[AI wiki page as an example]
This is the answer:
AI is intelligence demonstrated by machines.
This is the text:
[Some text you want to summarize]
This is the answer:

This allows more fine-tuning than just trying to reformulate the question you ask.

Also there's at least one marketplace for different neural network prompts including GPT-3, you can take a look there. Can't vouch for its quality, though.