How to Find the Gaps in a Discourse to Generate New Ideas (with a little help from GPT-3:)

Suppose you're interested in a certain subject and you would like to explore how you could contribute to the existing discourse. 

A good way to estimate where your contribution could be of value is to see if there are any structural gaps: parts of the existing discourse which could be connected but are not yet. This is the surest way to have a eureka moment: when you connect two ideas that are relevant to the discourse but have not been explored sufficiently in the same context, you will surely have a sort of epiphany, even if it is a small one. 

Below we demonstrate how you could do that using InfraNodus text network analysis tool that can identify gaps in a discourse.

The basic approach is to detect the gaps and to then ask interesting questions or to propose hypotheses that could bridge those gaps and lead to interesting ideas. You can do that by using the Insight Generation App in InfraNodus, which generates these questions with the help of OpenAI's GPT-3 natural language generation model. You can also make use of a GPT-3 based language generation tool, such as the OpenAI's Playground, Jarvis, or Copy.Ai to help you generate ideas based on those questions and hypotheses. Here's how it works. 

 

Stage 1: Get an Existing Discourse

First, you need to get an idea about the existing discourse on a certain topic. Suppose we're interested in martial arts. There are multiple sources to choose from:

• Google search results for the subject,
• research papers,
• Wikipedia articles
• related search queries (what people search for when they search for the "martial art")
• an authoritative book on the subject.
• our own text

You can use InfraNodus' internal import apps to get the data from any of the sources above just using the search query "martial arts": importing Google search results, a PDF of a book on the subject or simply copying and pasting a text of a Wikipedia article.

However, one really amazing source is the online dictionary Reverso.Context, which does not only translate, but also shows in which context a certain phrase is typically used. We love this data source as it's compact and impartial, giving you a quick overview of a topic without too much detail. 

Unfortunately, it does not have an API, but you can simply scrap the data manually (read this guide on scraping the data from the web).

 

Stage 2: Add the Data Into InfraNodus and Analyze the Results

Step 1: Copy and paste the source data into InfraNodus (e.g. the examples of the context where a certain topic or phrase is used). You can also simply start writing your own text. 

 

Step 2: The resulting semantic cloud will be visualized as a graph. InfraNodus will automatically identify which words tend to co-occur in the same context as "martial arts" using the data you provided — we call them "topical clusters". They will be shown closer on the graph and have the same color. Remove the words "martial" and "art" to see what topics are formed around them (read on Delete the Nodes / Words and Add them to the Stoplist)

Here's a link to this graph on InfraNodus, live:

https://infranodus.com/concepts/martial_art_reverso?background=default&most_influential=bc&maxnodes=150&labelsize=proportional&labelcolor=node&edgestype=line&drawedges=true&drawnodes=true&defaultlabelsize=14&labelsizeratio=2&dynamic=accumulate&cutgraph=1&selected=highlight&stopwords=art+martial

 

Step 3: Analyze the results. For instance, you'll see that when we talk about "martial arts", you have the 4 main topical groups:
• Japanese, ancient, Korean
• World Sport
• Developed, Aikido, Practice
• Chinese, Tai Chi

This gives us a very nice representation of the semantic field around the concept of "martial arts". 

We can use this information to explore the concept further. For instance, if we're interested in Aikido, we can make the same search on this topic and add the results to this graph (or make a copy of the graph and add the previous data + the new data on Aikido)

 

Stage 3: Generate New Ideas using the Structural Gaps in the Discourse

Step 4: Use the Insight panel to see what's missing from the current discourse: 

In our case, InfraNodus is proposing to link the idea of internal martial arts, like "Tai Chi" practised for health reasons and "Jiu Jitsu" and "Aikido" which were developed by the warrior class in Japan. 

We can ponder on the connection between these approaches and how they are different to each other. If we come up with an interesting idea, we could contribute to the existing discourse on the subject in an interesting way. 

...

Another option that you can use in InfraNodus is the Insight Recommender System, which generates a research question based on the structural gap:

You can also make the OpenAI's GPT-3 natural language generator come up with an idea for you. To do that, click the "Generate Questions" button and you'll see something like this:

If you like the questions, you can try to answer one of them to develop this discourse further. Alternatively, you can regenerate another set of questions or generate interesting ideas related to the structural gap.  

One of the ideas we got using OpenAI was:

"Aikido is a martial art that is based on the principle of meeting the aggressor's attack with a force that is not a direct counter-attack. With tai chi, the same principle of meeting an attack is used. Tai chi is also based on the principle that an attack can push you off balance in many different ways."

 

Why not? We can explore the idea of meeting an attack rather than counter-acting to it, proposing a possible relation between these different forms of martial arts. 

 

Step 5: Let's add this idea into the graph, and see what else InfraNodus has to offer. To do that, open the Interpret editor panel and copy and paste the phrase you generated above.

The Interpret panel will keep the original graph "clean" and let you create your own narrative on top of the existing one:

 

Step 6: Now, consult the Analytics > Insight panel on the right to see a proposition for a new idea, that is:

let's think of a connection between various physical movement forms, such as capoeira, which is more like dance without the physical contact, and the sports practices that have competitions and championships.

The recommendation received from the Insight Recommender System (based on OpenAI GPT-3): 

The connection is that they are all a way to express oneself, and to allow the body to be used as a tool.

we then developed this idea further adding some of our own thoughts on the subject:

While there are many different physical movement forms, like capoeira (no contact) and dance (which may have contact), as well as the martial arts that are also a form of sports (that have competitions and championships) — they are all a way to express oneself, to allow the body to be used as a tool. 

Add this into the graph and reiterate.

 

Step 7: Keep on connecting the different clusters of ideas to generate your original discourse on the topic.

In this case, we think of a phrase:

While all those martial arts are a form of expressing yourself, some of them are internal (so the expression occurs on the level of one's own experience) and some other forms are external (where your expression may be expressive, extravert and even designed for combat). Is there a martial art form that we can create that would combine all those elements?

The response we get from OpenAI: 

The internal arts are often combined with the external arts. For example, before you fight, you can do some sort of meditation or you can do some form of dance/movement to get you in the right mindset.

Let's add this into the graph, it's a pertinent idea.

 

Step 8: After a few iterations, we created a text on the martial arts, a proposition of a new idea where you practice not so only the physical movement, but also an attitude — towards yourself, towards the others and towards the environment. 

You can see the result in this live graph here:

https://infranodus.com/concepts/meta_martial_art_reverso?background=default&most_influential=bc&maxnodes=150&interpret=1&addcontext=martial_art_reverso&labelsize=proportional&labelcolor=node&edgestype=line&drawedges=true&drawnodes=true&defaultlabelsize=14&labelsizeratio=3&dynamic=accumulate&cutgraph=1&selected=highlight&stopwords=art+martial

The ideas you added will be shown in gray on the graph, so you can see how much of the existing discourse you managed to cover:

At this point, we can export the resulting text to edit the grammar and style and to post an article on this subject. 

In our case, we posted it on our website 8os.io, and made these ideas an integral part of our art project that links martial arts, dance, and technology. 

 

Stage 4: Add More Topics to Your Discourse

Step 9: We can use the same approach to elaborate further on this topic. 

For example, we can make a similar search for the topic of "attitude" to see what exists out there:

As we can see, it's a lot about changing one's attitude, so we could integrate this idea of personal change and transformation into our discourse on a physical / mental practice.

 

Step 10: Don't stop, keep on exploring the concepts and try out different sources. There is always something to add. Analyze your own texts to generate new ideas based on them. 

Feel free to reiterate and to even remove some statements and to generate structural gaps manually just from looking at the graph. You will see that the process is a lot of fun and the best part is that you see how the recommendation system is working under the hood, unlike so many black boxes today, actually revealing the whole process of thinking behind and making it more transparent and acting as a true collaborator rather than a mere assistant. 

 

Log on https://infranodus.com to try this approach on your own ideas!