WRT: Writer Response Theory

“ridcat” is a project that visualizes political speeches “from literary imagery to actual imagery,” producing a cloud of iconic photographs. The transformation is fascinating both in its products and through its process - a psychotherapy technique called Regression Imagery Analysis. WRT interviewed creator Neil Kandalgaonkar.

WRT: When did you develop ridcat, and when was it first publicly available?

Neil Kandalgaonkar: I hacked it up around September 2004, but only made a website for it around January 2005.

WRT: What inspired ridcat?

NK: Erik Frey, an acquaintance of mine, wrote some blog archiving software, called ljArchive. It had an unusual feature, “regressive imagery analysis”, that could categorize the content of your blog posts.

Regressive Imagery Analysis is a tool that some therapists use. The theory is that your thoughts will be revealed by word choice. So these psychologists tally up all the words from your therapy session, slot the words into categories, and then report that compared to the normal person you are precisely 28.2% more “oral”.

I don’t know if this theory is valid but it’s interesting, because you get a hard quantity for something qualitative. Erik’s software drew bar graphs. I thought it would be more vivid if I could turn those numbers into an image.

WRT: How does the ridcat visualization work? Did you experiment with any alternate methods on your way to arriving at the final look?

NK: It looks like a huddle of photographs, each iconic of whatever characteristic it’s supposed to represent. Sometimes that was easy — “aural” is an ear. Other times, I had to be more creative, to represent abstract concepts like “unknown” or “triumph” or “timelessness”.

Each photo is sized appropriately. If the text had a lot of “aural” words, then the ear gets correspondingly larger. I was adamant about making the photos cluster together, like a collage with a few dominant images, so this makes the layout very complex.

I didn’t experiment too much with different visualizations. I’m good at trying them out in my head first, so I usually have a vision of exactly what I want by the time I’m coding.

WRT: Did you originally intend to make ridcat a webservice that provided blog analysis? How did you end up focusing on selected political speeches?

NK: Yes, or just texts on the web generally. Unfortunately, my algorithm to do the layout is far too slow. I could definitely improve that, but I lost interest.

Why American politics — I suppose that just interests me, and I happened to be writing this around the same time as the big American political conventions were occurring.

WRT: You’ve commented several places that the Regressive Imagery Dictionary should be taken with a grain of salt - however it certainly does generate interesting results concerning, for example, the differences between characteristically Democratic and Republican rhetoric. What do we get when we use RID as a lens?

NK: One of the advantages of this technique is that it does cluster related words together, so some themes do emerge. Let’s look at these two analyses:

• 2004 Democratic National Convention
• 2004 Republican National Convention

Democratic speech is somewhat closer to the speech of everyday people. The size of an icon indicates how much it differs from the baseline. The Democrats’ icons are all smaller.

The clearest difference between the two is that only Democrats discuss pollution. However, the program mistakes that for Freudian ‘anality’. The Democrats also tend to emphasize the future, philosophical principles, and passing from one stage to the next. Perhaps this all adds up to a vision of moral progress.

The program believes that Republicans tend to emphasize themes of ‘restraint’. It’s partly because they harp on ‘tax’ a lot, as well as law and order themes. They discuss restraint from both sides — liberating voters from taxes, while promising to keep others under lock and key. Republicans also use ‘terror’ and ‘aggression’ images the most, so they contrast order with chaos, and their vigorous response to such threats.

But, I don’t know if any of the above is news to anyone. It is interesting to see a computer program extract such information automatically.

The one actually surprising result was that Barack Obama and Arnold Schwarzenegger are the only American politicians who talk about dreams. And Obama is the only politician at all who includes much positive imagery in his speech. Many people felt that Obama was a breath of fresh air, and I think this analysis helps clarify why.

WRT: Your use of a cloud of relatively scaled images is reminiscent of the phenomenon of tag clouds. Do you see any interesting comparisons or contrasts there?

NK: The tag cloud is supposed to give a quick overview the content of some large collection of media, indicating relative popularity of themes by the size of the word. That’s almost exactly like what ridcat does.

The differences:

• Tag clouds are usually employed in a system that distributes the tagging work to users. ridcat does it automatically. So ridcat’s inventory of ‘tags’ is rigid and it makes mistakes. Tags applied by users are, by definition, correct, and new tags are created all the time.
• I used a cluster layout, and tag clouds are typically alphabetical.
• ridcat is not tracking the absolute popularity of a given term, but what terms are more prominent than normal. So, for the most “average” speech, ridcat would just produce an empty diagram. (This also means that one could produce a ridcat-like diagram of themes that are unusually absent.)

WRT: You’ve described ridcat as moving “from literary imagery to actual imagery.” How does actual imagery work differently from textual analysis projects?

NK: Compared to other statistical analysis projects, it gives people a more visceral demonstration of what the analysis is saying. Rudy Giuliani’s speech at the 2004 Republican convention referred to terror or terrorism in almost every other word. In response, ridcat produces a diagram where the most prominent image is a screaming woman.

ridcat also has the advantage that it can group related words together and has a notion of the baseline of human speech. If you merely analyze the preponderance of words, you’ll either get trivialities like “the” or note that American politicians use the phrase “America” a lot.

No statistical analysis can reveal the essence of a text, since we have no techniques to comprehend the narrative let alone make connections to the real world. Emotional themes may be crudely extracted from a text, but why are they even there? It’s to drive the audience to a conclusion. The most important three words President Bush ever said are probably “axis of evil”. But he only said it once.

WRT: What kind of technology did you use for this project? Was there anything that worked exceptionally well, or anything that you would do differently?

NK: It’s written in Perl. My approach is not very good. I used a physics simulation — the images cluster like marbles in the bottom of a curved bowl, except these marbles are rectangular. This is very expensive to compute and Perl is very unsuited to that sort of task.

If I were to rewrite it, I’d use a faster layout heuristic. Then it might be fast enough to release on the web to analyze random texts.

The images were obtained from Stock.XCHNG. It was an exercise in incredible frustration trying to come up with appropriate photos, also freely licensed. This took more time than coding it in the first place.

WRT: What kind of future do you see for ridcat? Do you imagine it will ever be made available to users as a webservice, or released as open source?

NK: Probably no future, if it’s done by me. It’s fatally flawed since it hinges on word categorization from a disused branch of psychotherapy.

Coming up with photos to match abstract categories is hard. The layout takes too long, so it’s hard to make incremental improvements. The icons themselves are somewhat monotonous if you have spent as much time looking at them as I have. And, ultimately this is a very elaborate way of presenting a bar chart. It would be more interesting if there were extra dimensions to the data.

If someone wanted to take this up, I’d definitely help them. I don’t anticipate I’ll post the code; it’s pretty messy and hard to understand.

WRT: Any current interests or future projects you’d like to mention?

NK: I’d like to do more artwork with computation, but I’m busier with work these days. I do have some ideas.

Neil Kandalgaonkar is also the creator of 50 People See series of images. Each image was generated by “averaging” 50 photos publicly tagged with a given keyword (sunset, flower, eye, etc.).