Open dluman opened 6 years ago
I think I'm just trying to find ways to use Markov chain/LNN/RNN skeletons that I have lying around.
You might be interested in my own experiments with Markov chains. From the blog post:
Looking at the output of my toy Markov chain (“The window is slightly friendly”) makes me realize that the main problem I had with Markov chains isn’t necessarily the chain itself, but the corpus that people use to train on it. If you provide complicated prose to a Markov chain (such as Shakespeare’s plays), you are not going to get good results out of it. What I’m going to have to do is to simplify the prose to a level that a Markov chain will be able to grasp and imitate properly.
At the end of the blog post, I suspected that a simplified corpus might also work well for machine learning techniques as well. I never got around to actually testing it myself though.
I think the key reason for identifying a Markov-based method for this project is because I don't necessarily want the process to learn anything. Mostly, I just need to be able to reproduce the short language of ATC communications. The corpus will be a challenge, though, I realize. 30-day or younger transcripts are available largely as audio files through certain archiving sites. I may have to fold speech rec into this project, which is not an unknown quantity to me necessarily (I've done experiments with homophonic identification and espeak/Pocket Sphinx).
The assertion you make at the middle/end of the post of Markov chain-based processes typically creating nonce or absurd combinations unless given some guidance, is something that I am interested in actually going for here. There may be some bizarre contradictions that happen in the language.
The level of complexity is ratcheting up slowly on this one, meaning I have a lot of reading (and a little preparation) to during October.
One issue I need to address with this project is that I'm not interested in transcripts of airline tragedy. The audio archive at Live ATC provides somewhat current audio files (though I haven't listened to them to see what the samples sound like/how long they are), and I am thinking (because I love doing things that play on "current time" or concurrent time as much as possible) this will be a project that takes only files from one day prior, processing them and adding them to the corpus/chain to simulate a kind of development, a narrative richness that unfolds over time because the resource gets more and more diverse day-by-day.
one of the most riveting data-as-book works I've ever read. When crisis appears, it seems benign, but it's really troubling.
This has really got me thinking on a new idea that I'll probably aim for if my current one wraps up quickly enough. The idea of a computer-as-impartial-narrator seems to have a lot of merit, and presenting it as a data dump does a lot to excuse the clunkiness of trying to make a computer write like a human.
One of the issues we run across is that a reader takes a NaNoGenMo book and immediately starts looking for the "pattern" that makes it tick. Once you've read a few pages, you have a good idea of where it's going and how it was made, and then you're not motivated to finish it. I think one way around it is to simply make the real story happen "offscreen", and present a picture of it through the computer. The easiest compelling solution seems to be narrating details of a tragedy and letting the reader figure out what the tragedy was. That way, even after the pattern is understood, there is still more to discover through careful perusal.
Some ideas I'm kicking around now might be like -
or
And so on. There's a lot of potential here!
I really look forward to reading this when it's done, though. ATC chatter has always been fascinating because of the reasons above - a plane full of people headed somewhere, each with a different story in their own lives, and all summarized by a question about headings and landing angles.
In another installment of things that Doug wants to do, but we'll see if they actually happen, today's idea (I think I'm going to submit three) is to work with air traffic control (ATC) conversations to create a novel from the perspective of one ATC tower along a fictitious or real airline route. The thing that fascinated me most about flight simulators was always the ways in which pilots talk to each other and to ground. Whenever I used to travel frequently, I'd plug in my headphones and listen to the squawk between parties. Somehow, it relaxed me to fall asleep and wake up to the various hand-offs and brief conversations about weather.
This has a kind of literary precedent; I'm thinking of Seascape, by Heimrad Bäcker, which is one of the most riveting data-as-book works I've ever read. When crisis appears, it seems benign, but it's really troubling. I'll save the specifics in case anyone ever wants to read it. The physical book object is a fantastic work as well. Plus, there's a great essay by Charles Bernstein in it where he asserts that it can't possibly be a narrative or poetry, but I digress.
I think I'm just trying to find ways to use Markov chain/LNN/RNN skeletons that I have lying around. That seems to be one of the solutions here. However, I would like to make each page a daybook-as-map in which various planes and their messages are spatially located on the page to represent the origination/position of a given message. That seems like the complicating piece here that interests me the most. The unresolved question that I just haven't yet researched is where to find either current or historic logs of these conversations. I may also rely on flight simulators to get stems of what these conversations look like.
Another thing I thought of while writing the above is the complication that I'll need to store what other planes are doing. This makes me think of "Talk Town". So that'll be another wrinkle I need to smooth out for which I have some guide.
The title comes from the last message usually transmitted by ground to a plane, ending communication.
Let's see how crazy I can drive myself this November.