The Dowdbot challenge
« previous post | next post »
A few weeks ago, Maureen Dowd fantasized about a secret Google team trying to simulate her in software ("Dinosaur at the Gate", 4/14/2009):
When I ask [Eric Schmidt] if human editorial judgment still matters, he tries to reassure me: “We learned in working with newspapers that this balance between the newspaper writers and their editors is more subtle than we thought. It’s not reproducible by computers very easily.”
I feel better for a minute, until I realize that the only reason he knew that I wasn’t so easily replaceable is that Google had been looking into how to replace me.
There's a lot of far-out stuff over at Google Labs. But I'd be surprised to find that designing an army of Robot Maureens is in the mix, even though digital Dowd design poses some interesting challenges.
Among other things, any plausible imitation would need to master what the Dowd Report calls Dowdversions
The most lethal weapon in Maureen Dowd’s arsenal is the Dowdversion™ where two parallel phrases are used to create an ironic comparison often using a pun or play on words.
In order to emphasize the subtlety and complexity of the problem, I've decorated a few examples with green for non-parallel material, black italics for material repeated exactly, and subscripted red vs. blue to indicate words that are not identical but are parallel in sound, syntax, and/or sense:
Bush1 fancied2 himself the Decider3;
Obama1 fancies2 himself the Convener3.
W.1 lives in the shadow of his father’s presence2,
while
Obama1 lives in the shadow of his father’s absence2.
Even if Obama is treated as a superstar by
W.1– weary2 Europeans3,
some
Obama1-wary2 Americans3
may wonder what he’s doing there
Instead of
obtaining1 the girdle2 of the Amazon warrior queen Hippolyte3,
Obama has to
overcome1 the hurdle1 of the Amazon warrior queen Hillary3.
W.1 was2 immune3 to4 doubt
and afraid of it. (His fear of doubt led to the cooking of war intelligence.)
Obama1 is2 delighted3 by4 doubt.
Just as
Michael1 Bloomberg2 learned the perils of cuddling3 a groundhog4
when it bit him,
Mr.1 Obama2 learned the perils of coddling3 conservatives4.
They wanted1 him because he was the ultimate insider
and
they lost1 him because he was the ultimate insider.
Now Daschle’s punishment for
getting1 too_rich2 with special interests
will be to
get1 richer2 with special interests.
Sometimes the parallel items are in inverted order (chiasmus, in the traditional terminology of rhetoric):
We’re trading a dogmatic president for one who’s shopping for a dog.
The Dream Factory is being left in the dust by the Field of Dreams.
Sometimes the parallelism is merely the alliteration of onsets or other similarity of syllables:
Google is in a battle royal over whether it has the right to profit so profligately from newspaper content at a time when journalism is in such jeopardy.
The 53-year-old Schmidt is soft-spoken, exuding the calm knowingness of a therapist as he explains why privacy is passé and why passé newspapers are not going to pry more money out of Google to save themselves.
So here's a challenge to computational linguists everywhere: Can you create a Dowdversion Engine, along the lines of SCIgen , the Chomskybot, or the many title generators and band name generators out there? Bonus points if you can generate whole columns– but a sentence-at-a-time Dowdversion generator whose output is hard to distinguish from the real thing would be the most fun and interesting part of the project, I think.
Even a sentence-level Dowdbot looks like a hard problem to me, especially if we want the content to make the same sort of sense that the genuine article does. The author of the Dowd report tries his hand but (I think) falls short:
And that is the beauty of a well-played Dowdversion. It doesn’t have to make any sound sense, it just has to sound sensitive.
But surely the Regina George of political columnists is not trying to sound sensitive. The semi-parallelism make sound sense vs. sound sensitive may be a success in terms of sound, but it's a failure in terms of meaning — Dowd's shtick takes skill.
Anyhow, here's some more authentic raw material to get you started:
One seems small and inconsequential, even though he keeps insisting he’s not; the other grows large and impressive, filling Americans with cockeyed hope even as he warns them not to expect too much too soon.
Now it’s flyover country that’s starting high-flying trends.
Who can pay attention to the Oscar battle between “The Queen” and “Dreamgirls” when you’ve got a political battle between a Queen and a Dreamboy?
Instead of George Tenet teaching at Georgetown University, George Tenet should be taught at Georgetown University.
The brooding Scot ended his decade-long run as a hefty Heathcliff to Tony Blair’s chatty Cathy, stepping out of the shadows Friday with visible relief to begin a campaign for prime minister that he has already won.
Why can’t Google, which likes to see itself as a “Don’t Be Evil” benevolent force in society, just write us a big check for using our stories, so we can keep checks and balances alive and continue to provide the search engine with our stories?
Never has a star-crossed romance so perfectly illuminated a star-crossed conflict. The weekend meetings of the World Bank and the International Monetary Fund were consumed with the question of how the bank chief could fight corruption while indulging in cronyism. Who could focus on a weak yen when you had a weak Wolfie with a strong yen for Shaha?
In his first weeks padding around a White House that still has nails on the walls waiting for new pictures, and phone and e-mail kinks, Barack Obama could not locate the bully pulpit and ended up being bullied.
[As I should have remembered, there's at least one example of prior art, namely the 23/6 Dowd-o-tron. It's a good try, but (like the Chomskybot) it's too limited to be really interesting — and because of its "Chinese menu" architecture, it doesn't create any new Dowdversions… ]
Tyson Burghardt said,
May 21, 2009 @ 11:39 am
Numbers 2, 3, and 4 of the first examples seem pretty good. The first for profundity, the second for paronomasia, and the third for the audaciousness of the forced mythological reference.
The others I'm sure Google could simulate inside of a week of 20% time.
Aaron Davies said,
May 21, 2009 @ 12:06 pm
I see what you did there!
t.simeon said,
May 21, 2009 @ 12:17 pm
A Dowdbot would undoubtedly be a more adept plagiarizer. So many friends(i.e. crawlers) to put ideas in your head!
Ginger Yellow said,
May 21, 2009 @ 12:41 pm
"Dowdversion" seems to be a slightly artier version of the "In Soviet Russia…" gag.
[(myl) Good point — but Smirnoffisms always involve chiasmic role reversal, typically of the form "In [America or a substitute], you [verb group headed by X] Y; in [(Soviet) Russia or a substitute], Y [verb group headed by X] you", and are thus not so much less arty as more stereotyped.]
Benjamin Zimmer said,
May 21, 2009 @ 1:35 pm
Even more subtly, the Dowdbot would need to be able to mimic MoDo's ability to imply a chiasmus without explicitly laying it out, as in this example (which I discussed in a 2005 post on Dowdese):
Thomas Westgard said,
May 21, 2009 @ 1:50 pm
I'm switching standards here a little bit, but I think it's relevant anyway.
Creating a Dowdbot that replicates what Maureen Dowd might have written on a certain topic is interesting, but very difficult with current-day technology. What we're likely to see a lot sooner is a Reporterbot that gathers information and reduces it to a narrative format with its own flavor, which a certain audience might very much like to read. There are a lot of topics that are hardly covered, so there are audiences currently not served by any human reporter. Thus, any coverage might be an improvement in the eyes of those who care about that topic.
We already have Editorbot, in the form of Google News that parses through volumes of news, "decides" under an algorithm what's most relevant, and "prints" a page of news according to a preset format of how to make a page.
Reporterbot could equally well spider across the Web to find possible topics, "decide" under an algorithm what would please and audience, and "write" an article according to a preset formula of how to reduce the input to a story of given length and content.
A flesh-and-blood human will no doubt always be a part of how stories are found and told, but we will probably be seeing Reporterbots, if indeed they aren't already out there.
dr pepper said,
May 21, 2009 @ 5:50 pm
I recommend instead of a dowdbot, someone try making a dowderizer, a program that parses an existing piece of writing and creates dowdist versions of the sentences. It would be like certain gag translators that change wording into dialect stereotypes such as jive or redneck.
Ran Ari-Gur said,
May 21, 2009 @ 6:52 pm
@Thomas Westgard and dr pepper: You two should work together on this; you've already described a plausible two-module breakdown of the Dowdbot.
Nathan Myers said,
May 21, 2009 @ 10:53 pm
There are too many of her as it is. Wouldn't it be easier to make a program that writes better? Remember, it doesn't actually need to write well.
Chud said,
May 22, 2009 @ 10:45 am
Some of the "SAT Analogies" bits on Late Night with Conan O'Brien used similar techniques. An example (better than the Hippolyte example, though, of course, Conan isn't forced into prose) from nbc.com:
Alexander the Great: wept when there were no more lands to conquer::
Paris Hilton: slept when there were no more lads to bonk her