Hacker News new | past | comments | ask | show | jobs | submit login
Weird A.I. Yankovic: Generating Parody Lyrics (arxiv.org)
100 points by rbanffy 28 days ago | hide | past | favorite | 32 comments



I didn't catch that title at first, it's not Weird AL, it's Weird AI, as in Weird Artificial Intelligence.

From part of the conclusion:

"Indeed, one of the appeals of the Weird AI Yankovic, if there are in fact any, is seeing the train wreck of text that is output and deriving delight in the close-but-not-quite-sensical results. Weird AI Yankovic is a case study in design trade-offs—when to handle failure modes and when to permit failures—where improvements that reduce the likelihood of failed outputs also make the outputs “safer” but also less interesting."


> I didn't catch that title at first

For once, blind people using screen readers have an advantage. :)


For blind people reading this comment chain: upper case I looks like lower case l.


Hmm... honest question, do blind people learn the shapes of letters?

Surely even if blind from birth they'd at least want to touch embossed prints to feel the shape of characters? On the other hand I suppose there's not much of a need to memorize them?


Let's see if some dots will help.


Much improved. As would a different font, but that's like a religious argument that have been known to start Hatfield/McCoy incidents.


HN should adopt an IBM 3278 terminal font. ;-)


>> Lyrics parody swaps one set of words that accompany a melody with a new set of words, preserving the number of syllables per line and the rhyme scheme

That's an interesting definition of "lyrics parody" - interesting in that it ommits the most obvious component of lyrics parody, in that the replacement must result in a, well, parody. Of the original.

For example, take the result listed in table 1 in the paper, where the original lyrics of Micheal Jackson's "Beat It" are parodied by Weird Al Yankovic (with a lower-case "L") and the paper's system (Weird AI Yankovic, which an uppercase "i"). Jackson is first, Weird Al is second, Weird AI is third:

  They told him don't you ever
  come around her
  Don't want to see your face, you
  better disappear
  The fire's in their eyes and their
  words are really clear
  So beat it, just beat it

  How come  you're always such a
  fussy young man?
  Don't want no Captain Crunch,
  don't want no Raisin Bran
  Well, don't you know that other
  kids are starving in Japan?
  So eat it, just eat it

  The best part is that each taco
  contains a small
  To medium sized piece of sliced
  chicken nepal
  I don't think the food in question
  lasted awhile
  I promise, just promise
I'll spare the HN readership an erudite analysis of why Weird Al's parody is a parody- basically, it's funny and it makes one laugh. The Weird AI's "parody" is just a bunch of text that rhymes and sounds nonsensical, but it's missing the point of coming up with nonsensical rhymes; namely, to make you laugh. Or in other words not all nonsensical rhymes are funny and you can't call something a "parody" just because it's nonsensical and it rhymes.

But maybe this is a bit of a joke and I missunderstand it? I have been told that I miss a humour module, at times, so maybe it's my fault for misunderstanding why this system is interesting? "It generates nonsensical language that rhymes, surely that's an achievement", that sort of thing.

Oh, apparently the author contemplates a fart mode, following user suggestions. Maybe I really am misunderstanding this.

Maybe all of arxiv is really a big joke and I'm just not in on it.


> That's an interesting definition of "lyrics parody" - interesting in that it ommits the most obvious component of lyrics parody, in that the replacement must result in a, well, parody. Of the original.

All the GPT-3 stuff we've been seeing shows that generative models can generate text in the style of the inspiration text, with the topic of your choice. So that part "GPT-3, write me 300 words on the Amish people in the style of this text by Coolio" is kind of solved. But it wouldn't rhyme.

But now we have a generative text model that knows how to rhyme. Combine it with being on-topic and stylistically correct and get your "good" parody.


>> Combine it with being on-topic and stylistically correct and get your "good" parody.

OK, how do I do that? What do I do differently than the author of the paper?


Well you need GPT-3 access to start with. But I agree there's still an interesting paper to be written.


But, suppose I had GPT-3 access, what would I do?


> Maybe all of arxiv is really a big joke

So read it, just read it.


>>> Lyrics parody swaps one set of words that accompany a melody with a new set of words, preserving the number of syllables per line and the rhyme scheme

This is halfway like the Max Martin trick for songwriting. The apocryphal story is that this Swede who wanted to write pop songs in English didn't worry too much about the words and paid more attention to how they fit the melody. That's why the lyric to "I Want It That Way" parse a little funny. I know he's very intentional about syllables, but I'm not sure how much I believe the story--his English is actually really good.

Or you could say it's the Paul McCartney "Scrambled Eggs" approach. "Yesterday" might be the greatest parody ever.


Or maybe it's boring because it simply uses the established meaning in the context of music? https://en.wikipedia.org/wiki/Parody_music

In any case, you seem to be a bit conflicted on the meaning yourself, first demanding it meaningfully interact with the text of the original, and then relaxing the condition to the work only needing to be funny.

These are actually two rather orthogonal conditions: being funny or not is largely a matter subjective opinion, while interacting with the original text is something that can be more objectively decided by text analysis. And both conditions of course completely ignoring the musical aspect.


I'm sorry, where did I demand that "it" (the system?) meaningfully interact with the text? Also, where did I say it only needs to be funny? I can't find that in my comment above.

Regarding the meaning of parody, the link you provide continues as follows:

Although the intention of a musical parody may be humour (as in burlesque), it is the re-use of music that is the original[vague] defining feature.

In music, parody has been used for many different purposes and in various musical contexts: as a serious compositional technique, as an unsophisticated re-use of well-known melody to present new words, and as an intentionally humorous, even mocking, reworking of existing musical material, sometimes for satirical effect.

Examples of musical parody with completely serious intent include parody masses in the 16th century, and, in the 20th century, the use of folk tunes in popular song, and neo-classical works written for the concert hall, drawing on earlier styles. "Parody" in this serious sense continues to be a term in musicological use, existing alongside the more common use of the term to refer to parody for humorous effect by composers from Bach to Sondheim and performers from Spike Jones to "Weird Al" Yankovic.

So there are other uses of "musical parody" than to satirise (in Greek, "parody" means roughly "changed song") but the name of the system I'd say predisposes one to expect a humorous twist.

But of course I'm not surprised that I have to defend my interpretation of parody in the above paper. It is so, so confusing.

Edit: no, wait- so we agree that the parodies generated by the system in the article are not humorous? They are nonsensical and they rhyme, but they are not funny? In that case- why are they nonsensical in the first place? What is the point of a nonsensical parody if it is not humour? Am I really the one demanding a very strict definition of "musical parody", or is the author deliberately allowing a too-general definition, in order to make their system look like it's doing something it really, really isn't?

I apologise for revealing my deeply unfunny self in this thread.


No, I think you're on target.


I chuckled a little reading that, but obviously I'm not motivated to reproduce the results myself. So, maybe it's funny once?


You can play with the system in Google Colab.

https://github.com/markriedl/weirdai/blob/master/weird_ai.ip...

https://github.com/markriedl/weirdai

Not sure why the authors don't include these URLs in the paper. Working, open code is a major plus for a research paper. I had to track this down manually.


Note that, despite the title, this is not an article about Weird Al and his parody lyric writing process.


If you are looking for an excellent piece of journalism on that topic, I would recommend this article [0] from earlier this year. His approach to writing parody lyrics is quite interesting.

[0]: https://www.nytimes.com/2020/04/09/magazine/weird-al-yankovi...


You could be forgiven for missing that the title is "Weird AI" instead of "Weird Al" :D

For the disappointed, there was a nice NYT piece on Weird Al the human a few months ago, which had some details on his songwriting process: https://news.ycombinator.com/item?id=22861977


"Weird Al the human"

I wonder how long it'll be before that distinction needs to be made on a more regular basis.

"Barbara Walters the human not the Markov Chain base AI"


I read this as "Weird EY AYE Yankovich", which I thought was super duper clever.


how disappointing.


Example from the paper, alternate AI lyrics for Can't Touch This:

  Prompt: You cannot touch this
  And and and and should not touch this again
  Do not touch to the tongue
  The person in question
  And no one trying indigestion
  Do it in anger because they
  Should not touch the tongue say it ye and yea.
  I said unto you
  And if anything, uh, you pursue.


I laughed out loud when I got to do not touch the tongue.


Wise AI say

cat /dev/random > /dev/stdin

if [ self control < affection ]; then

        ProclaimAffection()
   else
        exit
fi


> serious paper about a silly task, written accordingly

I love it


Understanding humor is an important challenge in computation and language. If you are talking to who someone doesn't have humor, then you instantly know the person is probably stupid. The Eliza chatbot inside Emacs, M-x doctor, has a lot of humor.


> If you are talking to who someone doesn't have humor, then you instantly know the person is probably stupid

Okay but, keep in mind that just because someone doesn’t have the same humor that you do, doesn’t mean that they don’t have any humor.


That wasn't funny.




Guidelines | FAQ | Lists | API | Security | Legal | Apply to YC | Contact

Search: