NLTK uses statistical approach, which it seems he doesn't consider to be "NLP". TF-IDF is also part of that statistical approach. He seems to be talking about the old school rule-based natural language understanding.
If rule-based NLP is what he's truly referring to, he is actually very much on the money. Rule-based NLP gets too complicated too quickly and is useless when analyzing a language other than the target language or analyzing a piece of text that is poorly-formed (bad grammar/spelling) or heavily peppered with loan words and loan constructions.
Much like how Good Old-Fashioned AI has proven to be cumbersome and non-analogous to how biological brains actually work, rule-based NLP is flawed at its core. Statistical NLP has proven to be much more powerful, extensible, and tolerant to near-misses.
Summary: NLP is hard. There are lots of complicated academic concepts involved in parsing natural text that I don't understand. I have failed at using NLP in some indeterminable way. Please don't use NLP because you probably can't do anything with it either.
I totally agree with this, and the reason is simple: my US friends often fail to get the dripping sarcasm of my UK ones, so I don't expect a software package to.
From a practical perspective, it's remarkable how often you can cheat your way out of not needing NLP. Google is in a way a prime example of that.
In that list CYK, Earley, GLR, and Inside-Outside are taught as suitable for parsing natural languages (oversimplifying, the dividing line between parsers for computer languages and human language is that the former are deterministic and the latter spectacularly ambiguous).
So, you want to process a bunch of text, with time to parse each sentence increasing as a polynomial of the length of a sentence...I think you see where I'm going with this. Obviously, lots of people are looking for shortcuts here, but that probably means sacrificing something in accuracy, at which point you need to question whether you really want to be parsing, anyways. Last I checked, state of the art parsing accuracy is low 90s, and that's if all your sentences come from the Penn Treebank corpus.
So, when saying NLP is hard, NLP probably means parsing. You can probably describe applied NLP research as finding clever ways to answer the question you want to answer without actually parsing. Or parsing just a little bit, but not everything, you get the idea.
An interesting side question is just how do people parse natural language in real time so accurately. Or do people really parse at all, in the algorithmic sense?
> An interesting side question is just how do people parse natural language in real time so accurately.
We have hardware support for massive neural nets:) I would guess that as we read/listen we speculatively build several different parse trees that could represent the sentence, eliminating each one when its probability (based on the implied semantics) falls too low.
Just to clarify for everyone's benefit - guy rants about Natural Language Processing, and not Neuro Linguistic Programming, which is what NLP acronym is commonly used for and which is incidentally also quite relevant to the field of press release crafting.
14 comments
[ 3.1 ms ] story [ 46.2 ms ] threadit's not that hard; you just have to know when to use it.
Much like how Good Old-Fashioned AI has proven to be cumbersome and non-analogous to how biological brains actually work, rule-based NLP is flawed at its core. Statistical NLP has proven to be much more powerful, extensible, and tolerant to near-misses.
From a practical perspective, it's remarkable how often you can cheat your way out of not needing NLP. Google is in a way a prime example of that.
http://en.wikipedia.org/wiki/List_of_algorithms#Parsing
In that list CYK, Earley, GLR, and Inside-Outside are taught as suitable for parsing natural languages (oversimplifying, the dividing line between parsers for computer languages and human language is that the former are deterministic and the latter spectacularly ambiguous).
So, you want to process a bunch of text, with time to parse each sentence increasing as a polynomial of the length of a sentence...I think you see where I'm going with this. Obviously, lots of people are looking for shortcuts here, but that probably means sacrificing something in accuracy, at which point you need to question whether you really want to be parsing, anyways. Last I checked, state of the art parsing accuracy is low 90s, and that's if all your sentences come from the Penn Treebank corpus.
So, when saying NLP is hard, NLP probably means parsing. You can probably describe applied NLP research as finding clever ways to answer the question you want to answer without actually parsing. Or parsing just a little bit, but not everything, you get the idea.
An interesting side question is just how do people parse natural language in real time so accurately. Or do people really parse at all, in the algorithmic sense?
We have hardware support for massive neural nets:) I would guess that as we read/listen we speculatively build several different parse trees that could represent the sentence, eliminating each one when its probability (based on the implied semantics) falls too low.
Do you have any references to research arguing this, or is it just your personal idea?
http://en.wikipedia.org/wiki/Neuro-linguistic_programming