I wonder if, setting quality aside, the kinds of tricks (is "tricks" right?) that skaters do becomes more similar with technology. That is, skater X was going to do some random trick, until X learned that skaters Y and Z were doing quad jumps, so X figures he must do quad jumps now.
In other words, if we could represent skating routines as vectors, would the average cosine distance between all those vectors be increasing or decreasing?
Your comment is a bit ironic in the sense that I can tell you didn't read the article because you reproduce conclusions from the article. That's okay! Obviously you didn't need to read it to know what I would have said. :)
Let me quote from the end:
"Another argument against connecting distance and diversity is that distance is on a long running decline from 1900 even for the first four decades while diversity words were basically flat. When diversity words pop in the 90's there isn't an immediate reaction in cosine distance, it's only about a decade later, in 2000, that cosine distance takes a steep drop."
That seems awfully similar to the two points you've raised here.
What I do find a bit distasteful is that you jump in with "your diversity hypothesis is bunk" and accuse me of trying to fit a narrative - without even reading what you're commenting on.
For the title, I did say "At Harvard" and not "In Harvard" or "By Harvard". I think the student newspaper is indeed at Harvard. I'm also pretty clear about what text I'm looking at in the article. In the title I used "Harvard" over "The Crimson" because I figured fewer people would know "The Crimson" compared to "Harvard".
Regarding the flaws in the article (phew, I'm glad I didn't quite reach criminal level) - I'm curious which assumptions you think are horrible or what data shoddy. I don't think, for example, that I'm assuming the latent space model is "correct" as you say. I don't think I really have any significant assumptions about the technique or the meaning behind it. I read about the technique in the linked paper and reproduced it in my blog with a different dataset and found a similar result. It's strange to me that the signal produced by this technique is as consistent as it is across the 120 years of data. Beyond that, I'm pretty explicit that I don't know what it means or why it happens.
Regarding the "arbitrarily fit" line - as I say explicitly in the post, that's a regression plot to illustrate the trend.
Regarding the possibilities that The Crimson has more articles per year - it's true that's possible. It's not reality, they run about (for a generous definition of "about") the same number of articles every year. The articles do get longer over time. Either way, it's not clear to me what impact this should have on average cosine distance.
There are a lot of things that I looked at that didn't make it into the blog post. Without including them, then perhaps it looks like I'm cutting corners. If I did include them then I think the blog post would be shooting off in many directions. For example, I considered that political violence might be related - like maybe, in times where there's lots of political violence elite institutions come together and their language becomes more similar. That didn't really pan out though. I graphed a bunch of things that ultimately I decided didn't contribute very much and did not include.
Another way of thinking about it is in the original article Rasmussen (the original author) says "Look at this elite writing in NSF grants. The cosine distance is decreasing over time." I then say "Here is some elite writing - student newspaper at an elite school. Is the cosine distance decreasing there over time too?" And, it is. That's what the blog post is trying to say.
Now, maybe the latent space is "incorrect" - although Rasmussen and I use different embeddings that find a similar trend. Maybe it's not meaningful to use cosine distance in this context. But, it does seem like something has to cause it. Whatever it is and whatever it means, it doesn't look like the kind of thing that happens entirely by chance because it is consistent in different datasets and over many years.
One nice thing about getting feedback is learning all of the additional stuff I should have included in the blog post. I did look at number of articles per year, and it fluctuates, but there isn't a huge change across the century, and the change goes up and down. Total words, on the other hand, does trend up and goes up faster more recently.
I had a similar intuition and graphed unique words per year while writing this. I found, surprisingly, that actually the reverse was true. Unique words per year go up, even as diversity goes down. Another finding that may explain this is that the articles get longer as time goes on - so a simple unique word count may just increase as a function of the authors using more words. There is a period in the late 80's to early 90's where average word counts per article nearly double. I'd speculate that this is about the time The Crimson switched to using computers or good word processing or something that made writing articles easier.
A graph that may get to the heart of your question is something like "Unique word percentage over time" or maybe "What percentage of articles use unique words".
I suppose "old fashioned" is much more of a subjective measurement than an objective one. At first I was going to rate "snail" as not at all old fashioned, since we still use the word. But then, I thought of medieval snail drawings, and felt like the word has some element to it that does harken back to older days. "Citadel" is kind of a similar example. There is a modern hedge fund named Citadel and we still use that word, but it also has a connection to the past. I'd say Citadel is more "old timey" than "snail" even though both words are in modern parlance.
A better way to think about it is that these measurements are my subjective opinion on old timeyness and sillyness via an undefined and intuitive process for assigning values.
Thanks so much. You've really gone above and beyond with this comment. I very much appreciate it. I've made a couple tweaks that may resolve the issue. If you have a chance to take another look I'd love to know if it is or isn't working for you.
Hmm. That is odd, and not behavior I see in either Firefox or Chrome. I'm afraid my knowledge of CSS is insufficient to explain why that might be happening.
I was just trying to glibly acknowledge that some tweets are in a different language. I didn't filter those out. I do expect it will hinder the user's ability to tell which is offensive or not, but I don't think it will make much difference either way.
Were the images loading by default for you? I intended to set it up so that the images showed up as links for you to click if you wanted to or not - and there is a warning above the quiz about clicking the links. Although, granted, I could have been more explicit in the warning about what the images might contain.
I'm not surprised that it works poorly. I am surprised that it seems to work randomly though. I would understand if every Tweet that had a curse word in it was marked offensive, and that wouldn't be a very good filter, but it would at least be clear what was and wasn't "offensive" from Twitter's viewpoint. Here, there doesn't seem to be any way to tell.
The order of the questions is random per user, so not everyone will have the same 3rd set of Tweets.
I'm not sure if Twitter's algorithm is detecting a political bias and that detection is an input, or if Twitter marks Tweets as offensive because they have been reported by people who have a political bias, or if it is just random coincidence that one political side seems to get censored more than the other.
I think it depends on how you mean that. One way to imagine the phrase is to think about people with cloistered views in politics, religion, social views, etc. These people could get "offended" by being exposed to more cosmopolitan views but ultimately evolve to be better people by the experience - either by coming to better understand their own positions or by updating their beliefs to reflect things they have learned.
Another way to be offended though is just for jerks to say horrible things to you. Maybe that could be beneficial in terms of developing tougher skin or something, but it seems like that should at least be an optional thing. You might get tougher skin if I rubbed you with sandpaper for twenty minutes a day - but I should probably get your permission first.
I've seen tweets that are nothing but innocuous gifs get marked as offensive. Several examples in there are people responding to k-pop Tweets with identical statements (e.g. "I like so and so the best") and one will be marked offensive and the other won't. Sometimes almost identical tweets by the same person are considered offensive or not.
However sophisticated Twitter's algorithm is, and whatever data and behavior it takes into account, my contention is that it isn't very good and produces poor results. If people can't tell the difference between offensive and inoffensive tweets any better than chance - then what is Twitter really doing?
An earlier idea I had for this was to put two tweets side by side and ask the user to say which tweet was more offensive. Then, I figured I could use the ELO rating system to come up with an "Offensive score" for each tweet. e.g. "This tweet has an offensiveness of 2200" or something like that. I could then compare the average offensiveness of tweets that Twitter considered offensive versus not.
I wound up not going with that approach because many times you just have two completely innocuous tweets and picking which of the two of them is "more offensive" is just arbitrary. I could have curated the tweets so that only ones that were kind of offensive were in the quiz, but then I might be putting my thumb on the scales to get the answer I already believed in. I'd also need to get lots of ratings for each Tweet to have a stable score.
I think your idea is pretty interesting. It would allow a conclusion like "The average tweet Twitter marks as offensive is X% likely to offend a rater." I was coming at it more from a "Twitter's offensive identification is like random chance" perspective rather than just trying to assess quality. If I had considered this idea while creating the quiz I might have gone with it!
Hmm, that's unfortunate. I had considered loading the tweets as images instead of actually embedding them. I hadn't realized ad/tracking protectors might block the tweets, or else that might have pushed me over the edge to using images instead.
I think if you could genuinely identify offensive tweets then you could hide them from people who prefer that. It would make the user experience more pleasant. Few people want to be offended!
An ideal may be a customized algorithm per user. What offends me may not offend you. Of course, a trade off of that would be a growing filter-bubble...
That's true, but the end result, no matter how complicated their algorithm is, is that they don't seem to be doing a very good job of distinguishing offensive from inoffensive tweets.
In other words, if we could represent skating routines as vectors, would the average cosine distance between all those vectors be increasing or decreasing?