5 ms·

The reference at the end talking about comparing changes over time reminded me of a problem I've been kicking around. For this particular system that's not too

The reference at the end talking about comparing changes over time reminded me of a problem I've been kicking around. For this particular system that's not too hard, since you're putting each word on a one-dimensional femininity/masculinity scale, so you could plot a word on a line graph or something. But what do you do if you want to evaluate the changes of more complex relationships over time? Not just mapping words to a constant vector space, but modeling the relationships between words, such as clusters or word2vec representations. With something like word2vec you can take a bunch of words and project the vector space onto a plane so you can see the relative distances, but how do you express changes over time? You could show a bunch of planar projections for different instants in time, but it's hard to look at that and capture the changes. So how do you visualize changes to these more complex interactions between data points, and also how do you mathematically quantify some of these changes? I'd really appreciate any advice on this. And sorry that this is kind of off topic for the article :)

pg_is_a_butt 10y ago
measure change in relativity
joshuacc 10y ago
> With something like word2vec you can take a bunch of words and project the vector space onto a plane so you can see the relative distances, but how do you express changes over time? Animate it?
zebrafish 10y ago
I don't think I understand what your end goal is. what do you do if you want to evaluate the changes of more complex relationships over time? Which complex relationships are you talking about here? Edit: I think I understand what you're asking. Maybe one good way of doing this would be to cluster word groups for different periods of time. You could then perhaps look at where a word is now, where it was at the beginning of "time" and take the delta to examine words whose meaning has changed the most. Another approach might be to do apriori over time. Take a text from 1600 and lhs is "some manly action" and rhs is "he". Take a text from 2000 and lhs is "some formerly manly action" and rhs is still "he" but with much lower levels of support.
datalink 10y ago
One way is to learn clusters in the data over the whole time period and then calculate the cluster distribution of the same clusters for each time interval separately. You can then track the proportion of each cluster over time as a time series. So for example if you have a word cluster that describes computers, you could see it start growing in the seventies, while having near-zero proportion in 19th century etc.
jasonjayr 10y ago
The Gapminder charts very nicely illustrates multi dimensional data over time: http://www.gapminder.org/tools/#_locale_id=en;&chart-type=bubbles http://www.gapminder.org/tools/#_locale_id=en;&chart-type=bu... I imagine a dataset that charted the word use over time would fit into a chart like this.
viraptor 10y ago
What about 1d protection and simple line plots over time? You'll need 2 or 3 of them over some interesting projections, but it's better than unfamiliar and complex representation. Alternatively 1d protection + style change for another axis. Like changing the colour, thickness of the line, etc.
swf 10y ago
I did something like this in a project for my former work at a research group, and it was very difficult to visualize due to the number of variables we needed to communicate in the visualization. What we were trying to do was use word2vec to model changes in relationships between 700 proteins and one in particular, which is related to cancer/tumor growth. We created multiple word2vec models based on year windows of medical journals (so the 2003 model had input journals from 2000 - 2003). To visualize the models we used a D3 force graph where the nodes were the 700 proteins and the edges were known discoveries of relationships between proteins that had years associated with them - as in X protein was discovered to be related to Y protein in 2007. The relationship data was curated by people, independently of the word2vec models. The size of each of the nodes was determined by the word2vec model for the particular year's similarity score between that protein and the cancer-related protein we were interested in. To see the changes between years, we used a year slider which the force graph would respond to by animating the sizes of the nodes in line with the particular year model's similarity score for each protein. In addition, the color of the nodes represented changes in the models' similarities between the proteins - more green meant that the word2vec model's similarity score had increased compared to the previous year's model, and red meant it had decreased. The visualization is useful, but it is a bit of a mess considering there are 700 . I'll message you the link and if anyone else want the link I can send it, but I'd rather not post it here since it's hosted on my college's CS department server and it's not equipped to deal with a lot of traffic. Also if anyone has an idea of how else we might have done it I'd be interested to hear it. Edit: Didn't realize HN doesn't have a PM feature - if there's another way to send it to you and you're interested let me know.
- openasocket 10y ago
  Thanks, that's helpful! And you can email me at <my HN username> at gmail.com. I'm also curious if there are any ways to quantify, mathematically, the changes over time. There's the simple sum of the squares of the changes distance to get a sense of the "kinetic energy" of the system, but I'm wondering if there are some more clever analyses, especially something that can quantify localized changes versus global changes. Edit: so are you running a separate word2vec thing for each year's dataset? If so, how to you map between them, because the orientation the word2vec mapping generates will be random, and I worry that trying to rotate the mappings to some common axis could obscure some of the data.

pg_is_a_butt 10y ago

measure change in relativity

joshuacc 10y ago

> With something like word2vec you can take a bunch of words and project the vector space onto a plane so you can see the relative distances, but how do you express changes over time? Animate it?

zebrafish 10y ago

I don't think I understand what your end goal is. what do you do if you want to evaluate the changes of more complex relationships over time? Which complex relationships are you talking about here? Edit: I think I understand what you're asking. Maybe one good way of doing this would be to cluster word groups for different periods of time. You could then perhaps look at where a word is now, where it was at the beginning of "time" and take the delta to examine words whose meaning has changed the most. Another approach might be to do apriori over time. Take a text from 1600 and lhs is "some manly action" and rhs is "he". Take a text from 2000 and lhs is "some formerly manly action" and rhs is still "he" but with much lower levels of support.

datalink 10y ago

One way is to learn clusters in the data over the whole time period and then calculate the cluster distribution of the same clusters for each time interval separately. You can then track the proportion of each cluster over time as a time series. So for example if you have a word cluster that describes computers, you could see it start growing in the seventies, while having near-zero proportion in 19th century etc.

jasonjayr 10y ago

The Gapminder charts very nicely illustrates multi dimensional data over time: http://www.gapminder.org/tools/#_locale_id=en;&chart-type=bubbles http://www.gapminder.org/tools/#_locale_id=en;&chart-type=bu... I imagine a dataset that charted the word use over time would fit into a chart like this.

viraptor 10y ago

What about 1d protection and simple line plots over time? You'll need 2 or 3 of them over some interesting projections, but it's better than unfamiliar and complex representation. Alternatively 1d protection + style change for another axis. Like changing the colour, thickness of the line, etc.

swf 10y ago

I did something like this in a project for my former work at a research group, and it was very difficult to visualize due to the number of variables we needed to communicate in the visualization. What we were trying to do was use word2vec to model changes in relationships between 700 proteins and one in particular, which is related to cancer/tumor growth. We created multiple word2vec models based on year windows of medical journals (so the 2003 model had input journals from 2000 - 2003). To visualize the models we used a D3 force graph where the nodes were the 700 proteins and the edges were known discoveries of relationships between proteins that had years associated with them - as in X protein was discovered to be related to Y protein in 2007. The relationship data was curated by people, independently of the word2vec models. The size of each of the nodes was determined by the word2vec model for the particular year's similarity score between that protein and the cancer-related protein we were interested in. To see the changes between years, we used a year slider which the force graph would respond to by animating the sizes of the nodes in line with the particular year model's similarity score for each protein. In addition, the color of the nodes represented changes in the models' similarities between the proteins - more green meant that the word2vec model's similarity score had increased compared to the previous year's model, and red meant it had decreased. The visualization is useful, but it is a bit of a mess considering there are 700 . I'll message you the link and if anyone else want the link I can send it, but I'd rather not post it here since it's hosted on my college's CS department server and it's not equipped to deal with a lot of traffic. Also if anyone has an idea of how else we might have done it I'd be interested to hear it. Edit: Didn't realize HN doesn't have a PM feature - if there's another way to send it to you and you're interested let me know.

openasocket 10y ago
Thanks, that's helpful! And you can email me at <my HN username> at gmail.com. I'm also curious if there are any ways to quantify, mathematically, the changes over time. There's the simple sum of the squares of the changes distance to get a sense of the "kinetic energy" of the system, but I'm wondering if there are some more clever analyses, especially something that can quantify localized changes versus global changes. Edit: so are you running a separate word2vec thing for each year's dataset? If so, how to you map between them, because the orientation the word2vec mapping generates will be random, and I worry that trying to rotate the mappings to some common axis could obscure some of the data.