9 comments

[ 0.30 ms ] story [ 17.3 ms ] thread
It's interesting how you used the neural networks to overcome the vanishing gradient problem. The applications of neural networks are seemingly endless, as Sigmoidal.io demonstrates once again using it with natural language.
Clickbait. NLP is far from human-level accuracy and dwarfed by the success we have achieved in CV with deep learning.

If you consider language modeling the essential task of NLP, with deep learning, we can produce some interesting pattern, however still garbage anyway.

Completely agree. The examples are incredibly narrow and likely document and context dependent overall.
This is absolutely correct. Deep Learning techniques as applied to NLP have moved us forward -- but we are nowhere near human level
No doubts this is not true. We never solve the problem of understanding.
(comment deleted)
Well, accuracy depends on the specific use case, right? Ofc, in general, NLP is not at the human level, but the progress is astonishing. Gripping read.
Is NLP almost at human-level accuracy in picking up on subtext?

Can it formulate a human-level response to sarcasm? Could it respond in kind?

Does it know when a question is rhetorical? :)

This shallow article provides nothing to support its nonsense claim that NLP is "almost human-level accurate".

There exist tasks that can be solved at the level of human inter-annotator agreement, yes. That's how you know that the task is too simple or too factoid-driven, and we need a new one.