NLP right now looks like the computer vision 5 years ago: DL methods are starting to work really well, so a lot of "traditional" methods to process text might soon become obsolete.
The goal is to just feed gigabytes of raw text to a huge, complex neural network, and hope it will extract relevant features.
The problem is datasets. How can you distinguish a good result from a bad result? In some cases, depending on the user, it could be both at the same time.
Most advancements in ML is not accomplished by some new super algorithm. Rather, advancements are reached when new datasets are presented!
The goal is to just feed gigabytes of raw text to a huge, complex neural network, and hope it will extract relevant features.