Google Brain Researchers Make Significant Progress On Language Modeling (arxiv.org)
New submitter integralclosure writes: Using neural networks, Google Brain researchers have significantly improved a computer's ability to model English (achieving extremely low perplexity score on a large dataset). Using the model they were able to generate random sentences, such as the following: 'Yuri Zhirkov was in attendance at the Stamford Bridge at the start of the second half but neither Drogba nor Malouda was able to push on through the Barcelona defence.' The sentences are generally coherent and mostly grammatically correct. Advances seem to be a replay of neural networks' dominance in the Imagenet competition.
"achieving extremely low perplexity score on a large dataset"
That's quite an achievement, since I have a very high perplexity score on that sentence.
Do they check whether it achieves low perplexity only when the sentence is not, in fact, perplexing?
http://www.geoffreylandis.com
'Yuri Zhirkov was in attendance at the Stamford Bridge at the start of the second half but neither Drogba nor Malouda was able to push on through the Barcelona defence.'
Barcelona playing at Stamford Bridge? That hasn't happened for years!
They should hook this up to the Firehose and have it edit Slashdot stories.
F and U shouldn't be that perplexing than?
Time flies like an arrow. Fruit flies like a banana.
Where are we going and why are we in a handbasket?
Yuri Zhirkov was in attendance at the Stamford Bridge
it's just "Stamford Bridge", not "the Stamford Bridge".
Try again!
systemd is Roko's Basilisk.
Yes. Because it's not a simulacrum, it's a greatly simplified model. That it works at all suggests you've discovered at least a few of the important principles that let the vastly more complicated original work.
Also, you can poke at the model, see what improves it, what breaks it, plot receptive fields, all those things that are messy, difficult or unappreciated in an actual brain.
"The sentences are generally coherent and mostly grammatically correct," which means nobody will believe they are talking to a person.
I thought they said the program improved language modeling.