I trained a neural net to generate roller derby names & discovered it doesn't quite get the concept of puns. It's excruciating to watch it try.
@janellecshane La Vandastrophe isn't bad, though.
Cool! I wonder how you could provide context on puns? Word vectors might already carry some contextual information that might imply an unlikely conjunction, but probably one would have to first train a network to split a pun vector (??) into component word vectors?
Ooo, citation needed? :)
I have been wondering whether you could train a char-rnn to guess word vectors, so that when presented with a word outside a word vector set it guesses a vector based on what the word looks like. I suppose it would need information on context, POS tag, etc., but if it could work at all perhaps it could help unpack puns..
Here's the dataset (a bit messy though). https://www.kaggle.com/abhinavmoudgil95/short-jokes/
A word vector approach seems reasonable - maybe something like Robin Sloan used here https://www.robinsloan.com/voyages-in-sentence-space/
@cathal Here someone tried a word vector approach but the results weren't great. Maybe because the word vector was trained on politics. https://web.stanford.edu/class/cs224n/reports/2760332.pdf
@janellecshane LOL it'll come up with some real gems eventually though