Peterson Fellow Chronicles #3: Visualizing the "Rules For Life" Space

In this article, I show one way that we can visualize the "Rules For Life" design space that is created when we use standard Natural Language Processing (NLP) techniques on Jordan Peterson's original 40 rules for life.

What is a word embedding?

A word embedding is a list of numbers that represents a word or set of words.

To create a word embedding means to convert a word, sentence, paragraph, book, or any amount of any type of text into a list (or set) of numbers.

When we do this, we can easily compare our Rules For Life with each other.

One way that we compare them is through plotting each point in 2D or 3D. Each point might represent a word, sentence, book, whatever; in our case each point represents a single Rule For Life.

When we plot the points in a space that we can visualize, then we can directly see which Rules are close to each other in the space. The closeness of the points indicates the similarity of the rules. Similar rules tend to cluster around one another.

Visualizing the "Rules For Life" Space

I'd like to start by showing you all this 3D space, so you can see it with your own eyes.

Again, the distance between points indicates how similar they are to one another.

And again, each point represents a rule.

Try to notice any clustering of rules that you see. If you don't like awesome intro music, you can skip to about the 1 minute mark to just see the goods.

Creating Rule Clusters in the "Rules For Life" Space

What did you see? Based on what I saw I've decided on some clusters labels:

1. What to Do
2. What Not to Do
3. What to Be Careful of doing
4. When to focus on One thing at a time

I might pick them out like this:

Connecting the "Rules For Life" Space to Meta-Rules

These clusters appear indicate that some correspondence to our Meta-Rules from the last article (https://www.dhirubhai.net/pulse/peterson-fellow-chronicles-nlp-42-towards-meta-rules-life-loughnane/), which were:

1. Figure out what not to do first. Then aim at something.
2. Focus on one thing at a time, from the bottom up. Then you will know what to do next.
3. Focus on someone. First, make that someone yourself.
4. Get to know something. Develop yourself.

To investigate this, we can actually locate the Meta-Rules in the space and view them relative to the others 40 rules. I've done this below. Note that the diameter of each point indicates how close it is to you in 3D.

It appears that our Meta-Rules do a surprisingly good job summing up and spanning the "Rules For Life" space that we've defined.

Of course, when we insert these new points into the space, especially since they appear at the edges of known information, our perspective gets slightly distorted. I've attempted to label the clusters below so that they mimic the ones that I defined above in the original 40-rule space.

Observation Points-of-Interest

• It's challenging for these algorithms to tell the difference between "one" and "someone." Note that the One at a Time cluster seems to have gobbled up a number of new points (rules) in this new figure, versus the first clustering figure that I showed in this article.
• In reality, just thinking about it, the What to Do space actually spans both the Be Careful space, and the One at a Time space. In fact, contained within many of the What Not to Do rules are helpful hints about what to do (e.g., compare yourself to who you were yesterday, not to who someone else is today). In data science speak, these clusters are not independent of one another. In layman's terms, they all sort of "Venn Diagram" each other. In common sense terms, all of the Rules For Life are aimed at helping us figure out what to do.
• Keep in mind that these spaces each represent only one possible lens through which we can view our "Rules For Life" space. The exact way that a space appears to us in 2D or 3D is based quite a bit on our choice of methods and techniques. That said, the relationships between points within each space (i.e., how close or similar points/rules are to each other) should be relatively consistent across any spaces that we could generate. Said differently, these relationships between points are the patterns that we're looking for.
• Technically speaking from a data science perspective, the invariance of the relationship between points has manifested itself in the following way so far: no matter how I create the word embeddings, whether I use a simple Bag-Of-Words (BOW) approach or more sophisticated approaches like a Term-Frequency Inverse Document Frequency (TF-IDF) approach, I see the same similarities between words. The preliminary work that I've started to do with pre-trained embeddings points in the same direction as well. For the record, a TF-IDF approach was used to create the figures in this article.

High-Level Conclusions

1. The Meta-Rules seem to span the Rules For Life Space rather well, and provide a bit of a bounding box for them. This result was very encouraging.
2. It's not clear to me that a Be Careful space is all that useful, nor a One at a Time space. They were fun to play, though.
3. Further, it's not clear to me that the Meta-Rules map well to the clusters. Perhaps the Meta-Rules should be considered to be the centers of their own clusters. This needs to be investigated further.
4. The Rules For Life space seems to have some patterned characteristics that are insensitive to changes in simple word embedding creation techniques. Hopefully this continues as I throw more sophisticated techniques at the problem in the coming weeks.

What's Next

In my next article, I'm going to do a deep dive into a technical explanation of how I've been putting these analyses together, and I will also address the use of GloVe embeddings (https://nlp.stanford.edu/projects/glove/) that I mentioned during my 2nd article. I'm hopeful that by using pre-trained embeddings will be fruitful in my search to extract deeper meaning from Jordan Peterson's Rules For Life. I certainly believe that they will be an indispensable tools for me when I begin to analyze larger chunks of Peterson text, as that text will require a depth of linguistic/semantic similarity between words and phrases that is not achievable with the simple approaches that I've show so far.

I'm also going to make sure that my Python codes are posted on GitHub in easy-to-understand Jupyter-Notebooks (links forthcoming in next article), so you guys can all follow along.

One of my primary goals with this series is to keep most articles at a level that most people can read and enjoy. While my next piece will likely be a bit more technical, I'm working to balance worthwhile and meaningful conclusions and ideas with data science details. I will continue to aim at striking the right balance.

Lastly, always feel free to engage with my content and ask questions, leave comments, or provide other feedback that comes to mind! Thanks!

Cheers guys!