GPT-3 AI is best known for generating new content. However, it can also be used for interpreting what already exists and rewriting it in a new way. For example, generating a summary for a book or an article.
In the approach proposed below, as implemented in InfraNodus insight generation tool, we combine text network analysis and GPT-3 AI to produce highly relevant and original book and article summaries that will extend beyond the generic response that GPT-3 alone can provide.
The main difference is that instead of relying on AI to "understand" the main ideas in a text, we extract those ideas first using advanced language processing and text mining tools. We then generate a prompt using the main topical clusters and key concepts found in the text to produce highly relevant and original summaries that will touch upon all the essential topics in the original text.
One advantage of this approach is that it also works for bigger texts. For instance, GPT-3 is limited to about 16Kb per query (4000 tokens). Using InfraNodus, you can upload a whole book (e.g. 500 Kb), extract the most important topics, and then have a succinct summary generated for you. This methodology is akin to compression technology in that it reduces the dimensionality of text and, thus, the prompt required to generate relevant article summary.
Here is a step-by-step guide on how you can do the same:
1. Go to InfraNodus and create a new graph
Log on to https://infranodus.com/apps and then select Analyze an Article app (you can also use the Analyze a Book or the Summarize a Text apps).
2. Add your existing text
You can copy and paste your text or upload TXT / PDF / CSV / MD files.
For the sake of a demonstration, let's use a book called "The Conversation Book" by Circadian publishing house. It is a book of 40 questions that propose alternative hypothetical scenarios of life — to be played together with a partner in a conversation.
Upload this in InfraNodus and we get the following result:
3. Analyze the Content of a Book / Article
You will see a graph with the book's content. The nodes are the main concepts and co-occurrences of those concepts are represented with the edges. We then build a graph and apply powerful network science algorithms to identify which concepts have the most influence and which belong to the same topical clusters.
These topical clusters are then fed to GPT-3 AI and we generate categories for the book's content. In our case it's "Games", "Reality", "Conversation", "Life" — describing the content of the book really well.
4. Generate a summary of the content
We can then use the "AI Generate Summary" button in the Analytics panel to generate a summary based on the main topics identified as well as the most influential concepts:
As you can see, the result is very precise. It could be used as a text describing the product in an online store or to communicate the importance of the approach proposed in a book.
Here, we're talking about games and the importance of games to propose alternative versions of lives and to make mistakes in order to experiment and to learn from our experience.
5. Summarize the book
The summary provided above is based on the main concepts found in the book. If you would like the summary to be a bit closer to the actual content, you can use another feature of InfraNodus, which extracts the most important statements from the book and generates a summary based on them.
It is found in the Statements Menu > Essential & Summary:
As you can see, the result is exceptionally good.
6. Rewrite the content: create an article outline
Let's suppose you would like to write an article on the approach proposed by the book. Instead of generating a summary, you can generate an outline of an article based on the main ideas presented in this book.
To activate this, go to Analytics > Topics > AI Article Outline and GPT-3 will generate highly relevant article outline that relates to the main ideas in your discourse:
Feel free to try this approach on https://infranodus.com with your own content or books.