A couple years ago when Hurricane Milton was making its way towards Florida, I had a conversation with a friend about the different words used to describe incoming storms. Said friend raised the idea of plotting all these different words on some sort of semantic meaning matrix, comparing and contrasting their different connotations. Said friend also encouraged me to write about this on my site (fun fact, at the time my site was titled “The Blaugustsphere.” A far cry from today’s title!). I guess the idea has been riding with me in my head since then, because here we are now.
While I was initially going to write about storm words and heat words, really letting my Floridian self shine brightly, exploratory conversations revealed that many people already have strongly held beliefs about right and wrong in these realms, leaving me greatly uninspired. Material aging words, however – a much more natural connect to my whole schtick, and much less of a tapped vein in my opinion. Patina, distressing, fading, wear, rust, among others (stick around, you’ll see the rest soon!).
In order to push this thing to a certain personal standard of conceptual rigorousity, I worked with Stanford University’s GloVe (Global Vectors for Word Representation). It’s a bit complicated, but basically: GloVe is a machine learning algorithm trained on a huge amount of text in order to represent every word as a point in a 100-dimensional mathematical space, positioning each word based on its usage in context of every other word. Words like frog, toad, amphibian, lizard, all are near each other in this space. In addition, word meaning can be mapped based on distances between words. For example, the distance between “man” and “woman” is roughly equivalent to the distance between “king” and “queen.” This distance from one to the other can then be understood to capture that mathematical difference between the gendered terms.
I ran 16 different terms related to material aging through GloVe, specifically looking at how each term falls on 6 different meaning spectra. I think that ultimately I learned just as much about these different terms and how they relate to each other, as I did about the usage and limitations of this sort of linguistic modeling broadly.
An Aside: a light disclosure: The 100-dimensional space GloVe produces doesn’t inherently portray the neat and pretty picture I’ve depicted here. The reality is a lot more messy and linguistically inscrutable. In order to map the terms on the “meaning spectra” in a way that a human can understand, I had to select specific anchor words to construct the axes around. Now, with this technique (something that is a sort of standard when attempting meaning-mapping in this way), there is the ability to draw out these axes within GloVe, and it is fairly legit, but there is also bias in which anchor words I picked for each axis. This is all just to say: this isn’t some perfectly scientific assessment. There is a certain amount of unavoidable humanity in the mix. That said, we still are able to see something meaningful.
Let’s dig in!
Let me break down what I’ve got depicted here:
- each colored line is a different term that I’ve mapped — you can see which colored line is each word on the rightmost side.
- the parallel vertical lines depict various meaning spectra related to material aging. there are 6 total:
- environment/exposure <—> use/handling
whether the term is associated with exposure to natural conditions, or to usage - material-specific <—> material-agnostic
whether the term is associated with a specific material, or is used broadly regardless of material - structural <—> surface
whether the term is associated with the composition of the material, or just lies atop the surface - extreme <—> subtle
whether the term is associated with an intense change, or something less so - deliberate <—> accidental
whether the term is associated with intentionality, or happenstance - positive/desirable <—> negative/undesirable
you get the gist by now. whether the term is associated with “yay 🙂” or “boo 🙁”
- environment/exposure <—> use/handling
- The vertical axis maps how closely each term is associated with the far end of either scale, marked by a +/-5. The middle, 0, depicts a neutral association with either end.
- So, then, by placing a dot for each word on each parallel line, and then drawing a line through each dot, you get this beautiful mess of colorful spaghetti.
Below, I’ve broken down how each word fares across the 6 axes, and detailed what I think GloVe got right and what I think it got wrong.
Aged
Right off the bat with an interesting one here. It seems that GloVe gets this pretty wrong to me. In my opinion, “aged” happens as both a result of usage and environmental exposure. It is mostly material agnostic, in my opinion – it can describe cheese, whisky, wine, leather, denim, metal, etc. Things can look aged on the surface or be aged deep into the bones. I certainly wouldn’t say it’s fully “extreme” on the extreme <—> subtle scale. Things can be deliberately aged and it also happens as a matter of course. And while usually something being called “aged” is a sign of quality, I don’t think it’s so strongly associated with desirability to place it at a +5.
Clearly GloVe disagrees with me on these points, however. I think what is ultimately causing the wackiness here is that “aged” doesn’t tie so clearly to physical things. It’s most often outside the context of materials, instead being used alongside words like elderly, sickness, marriage, death (according to the TensorFlow Embedding Projector). This is very different from a word like patina for example, which has much less variability in usage. We are able to get some good insight into the limitations of our methodology by looking at how GloVe treats “aged,” noting that this methodology works best with words that are used within a specific context.
Bloom
This is more or less what I would expect, bar the last two axes. Bloom tends to happen to leather and chocolate specifically, usually as a result of the item being left in a particular environment (humidity, temperature fluctuations). It’s a merely surface-level effect, but is definitely visually noticeable. However, I certainly wouldn’t consider bloom to be something done deliberately, and it’s certainly not desirable. There are numerous internet threads on how to fix leather bloom, and lots of people questioning their bloomed chocolate. These last two axes are likely a result of bloom’s association with flora. I can’t imagine anyone’s ever been upset to see a flower blooming, and it’s not typically an accident when that happens — even in nature, it’s the result of things happening with some sort of purpose.
Broken-In
By and large I think this is correct. “Broken-in” usually refers specifically to textiles or leather, it refers to the composition of the item becoming softer or more pliable, it certainly isn’t subtle, it happens through the natural course of wear for the most part but also can be done deliberately (many people asking “how to break-in your boots” suggests the latter, and the answer being often “just wear them” suggests the former), and it’s seen as a good thing when a piece has become broken-in. However, I would personally say that “broken-in” is almost exclusively a result of use/handling, and not environmental exposure.
Burnished
A wonky one here, likely due to GloVe associating “burnished” equally with its material usage as its metaphorical usage in literature and journalism (“burnished reputation”). Burnishing is something that you do physically to a material, very specifically to metal or leather. It is also a much more surface-level thing, especially with leather. It’s subtle-ish, perhaps, but not so subtle to warrant an almost -5 on that axis. And after hand-burnishing 30 camera straps in a row, I’m extremely aware that that sort of thing doesn’t happen by accident. It should be more than just +0.5ish. GloVe did get the desirability correct, at least!
Corroded
About what you’d expect! Except, I tend to think of corrosion as being more extreme than subtle. And, certainly more negative than just a -0.5. Not much else to say here.
Crazing
One of the more accurate graphs. I would’ve expected it to land more squarely in the environmental arena, and crazing does associate specifically with ceramics/pottery (as far as I’m aware), but other than that GloVe did a pretty great job!
Distressed
I think GloVe has got this pretty accurate, but I’ll be curious to hear y’all’s thoughts on some of these. The first axis: in my mind, garments get distressed as a result of wear, usage, and not so much environmental exposure, but I could see it the other way. In addition, I think surface vs structure is very interesting here. Distressing typically refers to textiles, where damage to the surface is almost always some sort of tearing, cutting, ripping, which necessarily affects the sanctity of the weave structure. Distressing can also be things such as paint splatters, however, which are certainly just surface-level. In addition, I think it’s interesting that GloVe seems to be picking up on lots of distressing being done to garments intentionally nowadays, as opposed to happening through the natural course of wearing something. I would expect distressing to be around the middle of the deliberate <—> accidental axis.
Faded
To me, “faded” refers to a change in pigmentation, usually textiles (I’m thinking of raw denim specifically), but also paint, photographs, wallpaper. So I think GloVe got the material-agnostic part right. I think it got the first, third, and fifth axes wrong, however. On the first axis, while fading does happen due to sun exposure, I mostly think of it happening due to usage (again, in the context of raw denim). I can see it both ways though. On the third axis, I’m hard-pressed to see fading as anything more than surface-level. It’s merely a pigmentation change (right? correct me if I’m wrong). And, on the fifth axis, when it comes to raw denim at least, fading is certainly not purely accidental — there are more than enough guides out there on how to get the best fades for that. I think all the other faded-things tend to be purely accidental, so that mostly checks out. Axes two, four, and six all seem to be pretty on the nose though, at least in the denim context (I’m not sure that faded wallpaper is +5 desirable).
Patina
I would expect patina to land more in the middle on the environment <—> usage axis. It happens due to interaction with the environment, but the item is brought into the environment as a result of usage — wearing your boots on a hike, for example. Also, I don’t see patina as -2.5 material-agnostic. It’s almost entirely tied to either leather or metal, as I understand it. The rest of the axes though? Quite agreeable! Especially the deliberate quant — just as there are many instances of people asking how to best fade raw denim, there are many people asking how to get the best leather patina (the answer is of course, just as it is with raw denim, “just wear the damn thing!”).
Rust
Almost a nice, clean gradient. While rust might be just surface, as it advances it can affect the structural integrity of the object.
Seasoned
I think at this point I’m having a crisis about what is exposure vs usage. Cast iron gets seasoned because of its exposure to all the things that happen when cooking — oils, seasonings (unfortunately redundant), heat, a lack of soap. But, those things don’t happen on their own. They happen because the cast iron is getting handled by the chef (or in my case, the person struggling deeply to make a simple New York strip, wholly undeserving of the prestige of a chef title). Everything else here seems to map pretty neatly, though! It’s specific to cast iron; it’s surface level (evidenced by its easy undoing through some good ol’ soap-and-water); a well-seasoned cast iron is certainly not subtle; it does not happen by chance (unless you are in the habit of not washing your dishes, it takes a level of intentionality to properly clean your cast iron without compromising the seasoning); and, the whole draw of cast iron is the seasoning that happens and builds robust, deep flavors. I think that’s the draw, at least. I have one cast iron skillet and I hate it — I’m certainly no expert here.
Tarnish
Is tarnish really only a +0.5 extreme? Intuition says it’s quite a bit more obvious. Although, I guess that depends on how tarnished the thing is… it gets a bit circular there. And, does tarnish happen on purpose? Maybe GloVe is picking up on something similar to distressing, where it has become commonplace to speed-up the natural advances of wear and time to satisfy the desires of impatient consumers. I am glad that GloVe registers tarnish as more positive than negative — it’s giving me retribution for my parents’ ceaseless war against the inevitable tarnishing of their beautiful brass kitchen sink faucets.
Verdigris
Environment yes, material-specific yes (enough to be more than just +0.5ish, in my opinion), surface yes. But subtle? Have you seen our little old lady in the harbor, the Statue of Liberty? She wasn’t born green. In fact, her green-ness is perhaps her most famous trait next to her bigness (and probably the symbolism of it all, too). And deliberate? I’ve never heard of someone doing it deliberately. On the positive <—> negative axis, I’d guess a score much more neutral than it got. Personally, I quite like the look, and many antique collectors do too.
Water-stained
Certainly water-stained isn’t happening by hand nearly enough to place it at a 0. And certainly not intentionally enough to place it close to a +5. I think perhaps GloVe got this mixed up with “stained,” as in wood and furniture staining, which would make those two axes a bit more sensible.
Weathered
Is weathering deliberate? That’s my only gripe here.
Worn
An almost clean graph. The fifth axis does raise the question – things become worn (aged) because they are worn (donned). They’re not worn (donned) with the intention of becoming worn (aged), but also it’s not like they just become worn (aged) by chance.
So what did we learn from this? There are a lot of words related to material aging, and it’s a fun mental game to disambiguate them. GloVe gets a lot of things right, but it also gets stuff wrong. Semantic context mapping is finicky, especially with words that get used a lot outside of the specific arena you’re thinking within.
I decided to tally up how often terms were represented correctly and incorrectly for each axis (or whether it was unclear whether it was in/correct), in order to see whether there were patterns of concept contexts that were difficult to suss out with GloVe.
| Axis | Correct | Incorrect | Unclear |
|---|---|---|---|
| 1. usage vs exposure | 6 | 7 | 3 |
| 2. specific vs agnostic | 8 | 5 | 3 |
| 3. structure vs surface | 9 | 3 | 4 |
| 4. extreme vs subtle | 7 | 5 | 4 |
| 5. deliberate vs accidental | 5 | 7 | 4 |
| 6. positive vs negative | 9 | 5 | 2 |
Terms seem to most frequently get mapped incorrectly on the first and fifth axes. The differences between aging as a result of exposure vs usage are pretty nebulous, when the usage of a thing necessarily involves its exposure to the environment. Similarly with deliberate vs accidental, when things are used intentionally and the aging is a necessary outcome of that, what’s the proper understanding of the term on that axis? If we can’t clearly map out the ends of these axes, then we certainly can’t expect GloVe to get it right. After all, GloVe captures the way these words are used, messiness included.
Ultimately, oh ye faithful thinkers and doers: GloVe gives us a fun way to analyze words based on their relation to other words. It’s all very Jacques Derrida. As always, I encourage you to comment your thoughts, feedback, and criticisms! Ta ta~
p.s. Semantic Context Gradients hats and shirts available now
Leave a Reply