Listening to Topic Models

I want to explore alternate ways of ‘visualizing’ patterns in data, beyond the visual. To that end, I’ve taken the major topics & their proportions from a topic model generated with MALLET and run them through the Musical Algortihms site at EWU.

1. I obtained data from the Portable Antiquities Scheme related to ceramic building materials recovered by the scheme (why this and not something else? I’m thinking about brick these days. No other reason).

2. I created a topic model of the descriptor text.

3. I take the composition file that is outputed (the one that can be read as ‘in document 2 the major topic is 4 at 25%, then topic 6 at 12%…’ etc), and grab the topics and the amount by which they compose the document- so the first two columns. I turn the decimals into whole numbers by multiplying by 100.

4. I put these two columns into Musical Algorithmns. I perform the modulo scaling, then I invert the numbers. I used a 1 for the duration of the note.

You can listen to the output here

So what does it sound like? Well, I haven’t got there yet. But… if you do the whole process again, this time with topic models derived from writing qua writing (rather than database entries; the link takes you to topic models I did from posts on Play the Past), you get this.  Which sounds markedly different. More structure. Less repetition.

Anyway, this is obviously something that’ll require some more playing around (ha – see what I did there?)

3 thoughts on “Listening to Topic Models

  1. This is such a cool idea :D

    Have you concluded that Ceramic databases are not very good composers? :P

    1. Well, it’s interesting. You’d expect a database to be very structured, and for that structure to be reflected in repeating rythmns etc. There is a bit of a drone to it though…

      1. Well, datasets often have outliers and these outliers were actually noticeable in the music as well. I though that was pretty interesting.

Comments are closed.