Generating random text | Hacker News Reader

Generating random text

cs.bell-labs.com

95 pointswhiskers27 comments

Another practical example of Markov-chaining is http://www.x11r5.com/ - a robot that's various trained from IRC, Twitter and Identica content.

There's even a weekly podcast generated from news headlines: http://www.x11r5.com/radio/

I got an awesome one: “Wtf are you doing in 2008?" obama gave them to read it as a utility to basically throw random hacked kexts at the corner store stocks mexican coke.”

Also, personalized http://yes.thatcan.be/my/next/tweet/

I accidentally reinvented this algorithm many years ago as a C++ project. Here is an example of passages that it created:

sections of ice fell through the. invectives in which he had been wondering how. roman scales was in readiness. occasional murmur of pain that continued to torment. desavanchers dicksen dickson dochard du chaillu duncan durand was a. waging a war of extermination against. lively about it no snap or bite. chairs cane sofas carved wood pillars rose pillars. skirting an acclivity covered with woods and dotted with trees of very deep water. scratching as though he'd tear his nails out and sharp bite out. jerked by a sudden stoppage of the sled dogs barked. mentioned a cemetery on the south the four brilliants of the sky great and. ranks plunging into the flames would extinguish them beneath their mass and the rest were seen in numerous flocks hovering about the borders of some beautiful river until it fell. fridays well i hope that we shall overcome. emphatically an' i make free the. profitable the captains of the sea and consequently the balloon remained.

You can see more info about it or download my source code at:

http://experimentgarden.blogspot.com/2009/11/software-tool-f...

Markov chains are always good fun to play with... a few months ago I worked on a class project which generated markov models from Final Fantasy SNES tracks. (https://github.com/IsaacLewis/MidiMarkov). I should blog about it at some point.

I hadn't seen Shannon's algorithm before though, which looks a bit more memory efficient than the approach I used.

I remember reading about that in "The Information", where it is described how Claude Shannon did it.

Now I'll have to tweak my spam detection even more. Joke aside and somebody correct me if I'm wrong, spam probably runs on a simple wordlist type algorithm.

What is then the usefulness (I define usefulness extremely wide) of such a generator?

If you have a good generator for text, you have a useful language model that can be plugged into applications such as speech recognition, OCR, predictive text entry systems and compression.

Could you somehow use it in reverse? What I mean is, is it possible to get a random text generator for a certain language and then use it to determine, whether a given text is in that language or not?

Yes.

Given the string so far, see how probable it is that the generator would generate the next character, p(x_n | x_<n). Running through the whole string you can build up the log probability of the whole string: log p(x) = \sum_n log p(x_n | x_<n). Comparing the log probabilities under different models gives you a language classifier. For a first stab at the one-class problem, compare the log probability to what the model typically assigns to strings it randomly generates.

For more on information theory, modelling and inference you might like: http://www.inference.phy.cam.ac.uk/mackay/itila/book.html

I recently did exactly this to discriminate English text from gibberish.

https://github.com/rrenaud/Gibberish-Detector

It sounds like you are talking about a naive Bayesian classifier. PG wrote a couple of articles on his experience with these for spam filtering (http://www.paulgraham.com/spam.html and http://www.paulgraham.com/better.html). They're probably a decent high-level introduction to the area.

For a more in-depth, yet very accessible discussion, I would recommend "Speech and Language Processing" by Jurafsky & Martin (http://books.google.com/books/about/SPEECH_AND_LANGUAGE_PROC...). It's considered by many to be the Bible of NLP.

Bentley's collection of Programming Pearls radically changed my outlook (when I read it as a novice programmer, I considered programming itself to be an end, not a means to an end). I still read and re-read selections to this day, and I always seem to learn something new each time.

Would anyone care to share books of similar quality or importance?

Is it possible to generate text by a sort of reverse-LDA, where you have topics (per-sentence or per-paragraph, ideally) and estimate the probability of a word to appear in a given topic?

You could then use these topics to generate more realistic-looking text as this ostensibly wouldn't have the wild jumps from one topic to another that naive Markov chains have.

Has anyone done anything like this, or should I give it a shot?

That sounds possible, but gathering all the data for topic-specific training might be somewhat maddening. The problem you get with larger groupings of words is lack of cohesion. If you trained it at a sentence level, you might be able to produce coherent sentences. But as you generated more sentences to produce a paragraph, it would likely meander.

I created a kind of Mad Lib generator using CFGs. A paragraph consisted of: [Intro] [Supporting sentence 1] [Supporting sentence 2] [Supporting sentence 3] ... [Conclusion]. All the sentences had placeholders for various nouns and adjectives that could later be filled in programmatically, and I extended the grammar spec to both permute sets of sentences and generate null productions with certain probabilities.

The base sentences were created by humans, about 15 per grammar rule. A single person could create a topic-based paragraph/grammar in less than two work days. The chances of it creating the same template twice was about one in a billion. Of course, the probability varied depending on how many seed sentences were present.

If the person writing the seed sentences is literate and passed the 6th grade, then everything the program generates is indistinguishable from human text.

It works marvelously.

That's interesting, do you have any code or examples?

My thought about topic detection is to have it learn which words go together, and then augment the Markov chain model by some method that would weigh the Markov chain probability with the topic probability to select the next word, so it would at least generally stick to topic-relevant words. Perhaps you could even select one topic (a sample sentence, really) in the beginning and have it generate sentences based on that for the entire document.

I wish I could publish it, but my company isn't very much into open source. It's a standard context free grammar framework modified to generate output in a stochastic manner. So it's basically a stochastic context free grammar (SCFG). I can go more into depth in private if you like.

The phrase for finding word pairs in text corpora is "cohort analysis". I was a on research team that did studies of that; mostly finding them, not generating anything with them.

It's an interesting subject area.

That gives me a good idea for further research, thank you.

Natural language generation is an established idea, so you should definitely hit Google Scholar before trying anything.

I will, thanks!

"Dive into Python" contains chapter on XML processing[1] that uses grammar defined in XML to generate random English passages. A very funny and informative reading.

[1] http://diveintopython.org/xml_processing/

Markov Chains are really handy. You can also use the Hidden Markov Model to do voice recognition. http://en.wikipedia.org/wiki/Speech_recognition

Shameless Plug: This is very similar to what I use in http://wordum.net/ but instead of letters, I use whole words to generate the text.

That's exactly the idea that popped in my mind when I started reading the article (except I would have done it for free).

How successful is it? What are the advantages compared to eg randomtextgenerator.com?

I did this exercise on my blog at http://programmingpraxis.com/2009/02/27/mark-v-shaney/.

http://www.jwz.org/dadadodo/ DadaDodo is a fairly old implementation that may be of interest.

Weren't they looking for help data mining the Palin emails? That'd be a better selection of random gibberish wouldn't it?

Please, no, just don't.

Reply on news.ycombinator.com