3 ms·
1) Deep Learning-based techniques outperform hand engineered natural language processing stacks in every measurable way. 2) There will be considerable progress
by nl 11y ago
1) Deep Learning-based techniques outperform hand engineered natural language processing stacks in every measurable way.
2) There will be considerable progress made on the Winograd Schema challenge.
3) The (unconstrained) Turing test won't be "officially" passed until October 2017 though (it will actually be passed in July or August, but the news won't leak until October)[1].
4) Some people will continue to insist that Deep Learning is nothing different to what was being done in the 1990s. At some point someone will get frustrated enough with this to blog everything that is different now.
[1] Specific enough prediction?
- jrk_ 11y agoWhy are you so specific about 3)? If it's a joke, then sorry, I don't get it. :)
- sushirain 11y agoWhat do you mean by hand engineered? Is a part-of-speech tagger trained on a manually annotated corpus hand engineered? If not, then which "stacks" are hand-engineered and still popular today?
- nl 11y agoYes this is true. Perhaps I should better express it as a hand engineered feature pipeline, ie the way people these days will POS tag, then lemmatize, then do something with WordNet, then maybe some Semantic Frame thing and finally have an application they can benchmark. Writing that pipeline is manual, and it has hand engineered features in it too, though some may be trained.