For my exercise I wanted to figure out if I could attempt a topic modeling of a film. Obviously for the modeling to work, the film would have to be translated into a text file. I thought of using either the screenplay or the close captioning subtitles as the document to input. I opted for the close captioning since they represent a kind of transcript of the dialogue and basic actions of the film, though I am open to discussing whether this was a better choice (and may end up doing a comparative topic modeling later to see what kind of results come up). 

The film I chose was Children of Men, in part to justify my use of the close captioning since it is a film that features a lot of expository dialogue (even if it's often from spoken advertisements). 

One of the things remarked by the readings was that choosing the number of topics to extract from a corpus is more of an art than a science, so I decided to play with this variable and run several tries for each number of topics. Here are some results:

TABLE 1 -- Topics:10 Try:1 Ordered from most to least prevalent

TABLE 2 -- Topics:10 Try:2 Ordered from most to least prevalent

TABLE 3 -- Topics:5 Try:1 Ordered from most to least prevalent

TABLE 4 -- Topics:5 Try:3 Ordered from most to least prevalent

*I established prevalence order by the number of documents in which that particular topic had the highest probability (i.e. 1 had the highest probability in the most number of documents)*

I'm trying to make sense of what kinds of variation occur given the number of topics and times the model is run. For now let me focus on the topic word immigrants, which is a main theme in the film and as such, along with the main characters' names (Theo, Julian, and Kee), should appear in all instances. Two peculiarities from looking at this instance: 

1. Words are grouped together into topics by how often they would appear within the same document, so it's interesting to note how varied the associated words with "immigrants" are in just these four instances. The one exception seems to be that in most of these cases (and in others I didn't picture) "immigrants" shows up in the same topic as "fuck" or "fucking". 
2. Given that topic modeling depends on probability of finding a particular set of topic words within a corpus, it seems a bit strange that in Table 3, the topic that includes "immigrants", "diego", and "road" is the most prevalent while in Table 2, the topic that includes these same words is near the bottom.