An Anthology of Fake Speeches

[BrianHicks]

I haven't done a whole lot of generative text, but I've got a couple ideas:

• generate a world and then tell a story in it, random-walk style. Kinda LoTR in Dwarf Fortress.
• make a book of poems (because making something interesting that rhymes sounds like an interesting challenge)
• make a book of fake speeches. There are plenty of speeches online, slicing them up and recombining in interesting ways might work!

That said, I'm not too sure about the domain here, so if any of those sound easier or harder for someone with little computational linguistics experience to tackle, I'd love the feedback.

[ikarth]

Those are all good ideas. I'd guess that how easy they are depends on which aspect of text generation you want to dive into. I'd rate them as going from hardest to easiest, though that's partially colored by my biases:

• Making a book of speeches is probably the easiest, depending on how you go about it. A basic Markov chain randomization and some interesting source text is the most straightforward way, but there are an infinite number of ways you can approach this. Lots of unexplored ideas for getting it to make sense.
• Medium difficulty is the poems. There are libraries that can help with detecting meter and rhyme, of various levels of accuracy; check the resource threads for a few of them. Or you can write your own if that's the part that interests you, since it's not a perfectly solved problem.
• I'm personally biased towards the simulation approach at the moment...mostly because I haven't successfully gotten it to work yet. So that's probably the hardest, though it likely varies based on the quality of the prose you're aiming for. Also, this is probably going to be the approach that's most concerned with the larger structure of the text and least concerned about the fiddly linguistic details. Though again, that'll very based on how you design it.

Either way, looking forward to seeing what you come up with.

[cpressey]

I like the speeches idea, probably because speeches are heavy on rhetoric. It would be interesting to see how much (if any) of that "oratory voice" remains after combining multiple speeches (even ones on vastly different topics) in some way.

[enkiv2]

There have been prior examples of trying to generate speeches that have been fairly successful. I specifically recall someone using char-rnn with all of Obama's speeches, and again with TED talks; the TED talks ended up being less coherent in terms of topic but otherwise very stylisitically similar to normal TED talks, and comparable in the amount of content to the low end of TED talks: https://medium.com/@samim/ted-rnn-machine-generated-ted-talks-3dd682b894c0

This indicates that, with a larger corpus (say, the set of all state of the union addresses -- which are in the public domain & available from several sources), you could probably produce significantly more complex & interesting speeches.

On Fri, Oct 30, 2015 at 7:53 AM Chris Pressey notifications@github.com wrote:

I like the speeches idea, probably because speeches are heavy on rhetoric. It would be interesting to how much (if any) of that "oratory voice" remains after combining multiple speeches (even ones on vastly different topics) in some way.

— Reply to this email directly or view it on GitHub https://github.com/dariusk/NaNoGenMo-2015/issues/46#issuecomment-152505133 .

[ikarth]

Not to mention you can throw in, say, Mark Twain's speeches if you're looking for a less serious result.

There is, on the other hand, a certain purity to the idea of the Ur-State-of-the-Union speech.

[BrianHicks]

Thanks all, I'm going to give the fake speeches idea a shot! I'll update the title of my issue.