Published on June 22, 2025

Word Chains

When I look up a word in the dictionary, I get unusually annoyed when I see the word itself used in its definition.

redundancy /rĭ-dŭn′dən-sē/ noun

  • The state of being redundant.
  • Something redundant or excessive; a superfluity.
  • Repetition of linguistic information inherent in the structure of a language…

The American Heritage® Dictionary of the English Language, 5th Edition

Yeah, ok, it’s not the exact same word and the definition makes sense, but I’d rather not have to look up the definition of “redundant” to understand what the “state of being redundant” is. The 3rd definition given above gets it.

So that got me thinking, how often does that happen and does it form a chain of words? Let’s define every word in the dictionary as a node and draw lines to every word that appears in its definition.

For simplicity I:

Some of the filtered words include Salpiglossis sinuata, sea snake, polling booth, occipital protuberance and my favourite, beefsteak geranium.

There were 67,780 such entries out of a total of 161,705. Initially I was also getting errors for around 12k words but that’s because the .words() method was returning all adjectives tagged as a but their internal tag elsewhere in the wordnet was s.

As for the actual word chains, I just straight up brute-forced it. No fancy data structures or clever algorithms. I just looped over the word: definition map twice and checked for matches. Yeah, it’s dumb, but it’s was a Sunday morning. It took about 2hrs to process on my M1. Let’s plot the results!

Ah yes, as expected.

I am using the vis.js library for this (first time). Whilst the code is working, the visualisation is simply too dense. I needed to do some more filtering. I removed words with 0 chains and made sure I am not double counting things. Oh and I need to turn on the physics setting!

Let’s check out how things look on a random sample of ~1000 words.

Note: it’s interactive! 🕹️

0

Great! We’re starting to see some clusters emerge! The arrows indicate who appears in whose definition:

twisting âźą worm


This means “twisting” appears in the definition of “worm”.

I pre-computed the above graph so that it loads faster, which will come in handy for the full results below.

A month later: optimising, hacking & crying

Future Petar here. On that beautiful Sunday morning everything indicated that I was going to finish this before noon, just in time to go out and enjoy a nice day at the beach. What a fool I was. Apparently, visualising 85441 nodes and 384717 edges is more computationally expensive than landing on the moon. I spent a few more Sundays trying to render this giant network…

TLDR

❌ What didn’t work:

I wish the batched approach worked because it resulted in some cool behaviour. See the below (sped up) video:


âś… What (kind of) worked:

Unfortunately this approach also immediately starts to chug, even with just ~7% of all nodes and ~3% of all edges, but if you wait ~2 minutes you get a semi-usable network. Interactivity is choppy but, hey, it works.

Final results

Here’s the precomputed network for 5839 most-connected words (10000 edges) after a few minutes of stabilisation. Give it a sec to load. This one is also interactive, but you’ll notice it’s not as “floaty” as the sample earlier; that’s because physics have been turned off.


I am disappointed I didn’t get to load the entire network but I am kind of tired (and bored) of this side project – I first jotted this idea down back in 2021 – better to publish it now rather than never. With a bit more tinkering I think I could get this to work and also look into some questions such as “what’s the longest chain?”. If anyone has any ideas on how to improve the performance, here’s the Github repo. Thanks for reading!