The NLP progresses seems to me that has split in two big group of thoughts:

  1. Deep learning have a neural network with multiple layers that has been trained and learned to represent knowledge and understand the meaning of sentences. However, such system would be difficult to debug and correct

  2. Logic and inference, label parts of speech and learn by inference and using tools such as Natural logic.

Can someone point out if anything is wrong in these conclusions? Is there any other stream of thoughts that I should be taking into consideration?

  • 1
    $\begingroup$ Which conclusions? Note that your last question effectively asks for a summary of the field, which is too broad for SE. $\endgroup$
    – Raphael
    Dec 31, 2014 at 15:44
  • $\begingroup$ I am sorry if it sounds very broad. I am asking for correctedness for my statements and if I am missing something on the field $\endgroup$ Dec 31, 2014 at 16:54
  • $\begingroup$ Note that many machine learning techniques are employed in NLP, not just Neural Networks. For example, SVMs are very effective at Word Sense Disambiguation. $\endgroup$ Dec 31, 2014 at 23:28

1 Answer 1


Very briefly:

Option 1 faces difficulties since certain connectionist models are incapable of learning natural languages.

Option 2 faces grave difficulties due to many negative results, notably Gold's theorem. There's a nice overview which includes a discussion of probabilistic approaches.

Generally, the response to these difficulties has been to look at other sources of information beyond the linguistic input alone.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.