Hacker News Reader

Comment by caiobegotti | Hacker News Reader

As a linguist and software engineer I can't imagine someone doing serious NLP without ever having studied [concrete] syntax trees and such. It is easy to impress people with some tokenization but it's n-grams that are really useful in the real world, as is understanding syntax trees and all the interconnections possible inside them so you can NLP the shit out of real world text/speech, instead of simple examples with a tagger (and a good, carefully crafted for a demo like Apple's, trained set of tagged data). This is a good summary article with very good links nonetheless.

View on HN

As a linguist, I see what you mean. However, I disagree on a couple of things. First, I wouldn’t say “n-grams and understanding syntax trees and all the interconnections possible inside them” will make your NLP skills or results any better. Understanding n-grams and syntax trees should be quite easy for linguists, but that won’t do the trick. It turns out that tagging, lemmatizing, wsd, parsing, and many other fundamental tasks in NLP are not that “simple” after all. Of course, you can use libraries to do all of those tasks -and this is pretty simple, indeed,- but most libraries will end up making flagrant mistakes if texts -either spoken or written- get too complex. Second, real world texts are complex by default, so, for someone to “NLP the shit out of real world text/speech”, they will have to find creative solutions to make tools improve -and that can get pretty difficult. Once again, understanding n-grams and syntax trees is not enough. There’s one thing I totally agree with you on: tokenization, tagging, and the like might not be that impressive. Nonetheless, creative solutions to the problems underlying those tasks are, in fact, rare and impressive.

> as is understanding syntax trees and all the interconnections possible inside them so you can NLP the shit out of real world text/speech

But do modern DL approaches (e.g. SQuad, translation models) defy this approach? They train DL models on labeled data without knowing anything about syntax trees, and allow NN do all the magic..

That's how NLP can be a dead end by putting your faith on NN.

So far this dead end outperforms traditional ways though.

Lol for sure any NN-based approach can claim to outperform traditional ways while half of CVPR paper results cannot be reproduced.

CVPR?

Conference on Computer Vision and Pattern Recognition

> while half of CVPR paper results cannot be reproduced.

even if just another half of results are proven to outperform traditional approaches, is it strong enough trend?

You clearly did not see that was sarcastic.

From reading papers on WaveNet and AlphaGo, the problem seems to be incomplete information in the papers, but I am not going to argue with the popular results.

You could jointly train a NN to parse the sentence and tackle the NLP task you have at the same time, or instead of just using the tokens you could use a parser to add more features too.

edit: clarification

> They train DL models on labeled data without knowing anything about syntax trees, and allow NN do all the magic.

Sure, but that doesn't mean that knowing about syntax won't improve the result. If you're training a translation model on a huge database of labeled examples, it might discover syntactic relationships from scratch. But if you don't have so much data, you're probably better off using all the auxiliary information you can get.

> But that doesn't mean that knowing about syntax won't improve the result

this is likely correct, additional high quality input will likely improve performance, but creating such input for 200 modern human languages requires significant effort, much larger than allowing NN to solve this problem, that's why researchers invest this effort into NN improvement, not syntax tree creation tooling.

All of us have to strike the right balance between stubbornly believing that NNs will solve all problems in NLP -which, imho, underestimates what linguistics can do for this field of study- and stubbornly saying that only linguistic knowledge will lead to an improvement of the results -which, imho, means underestimating the outstanding work on deep learning techniques done so far. As you said, creating high-quality input is likely to be costly -and I'm not talking about syntax tree creation tooling only here. But, isn't it worth the while? I believe it is.

Sorry, but for a lot of situations n-grams just don't scale. There are just way too many combinations. There are a lot of people dealing with GBs and TBs of text.

> "There are just way too many combinations"

truncated SVD

Nowadays, deep learning powers most of NLP.

The state of the art is primarily hybrid models which use deep neural networks in combination with classical parsing algorithms. For example, the Stack LSTM method uses a form of arc-std with an LSTM and a stack of sorts.

And on which problem this approach achieved state of art results?

Dependency Parsing [1].

[1] https://arxiv.org/abs/1505.08075

I checked this paper briefly, and couldn't find any "classical parsing algorithms" there. Maybe you could point me on what do you mean?..

Also why do you think it is state of art? Is it in top results in any recent benchmarks (e.g. http://universaldependencies.org/conll17/results.html)?

Arc-standard [1] is the classical algorithm in question. See section 3.2 of Dyer's paper: "Our parser is based on the arc-standard transition inventory (Nivre, 2004), given in Figure 3".

Perhaps it's not still state of the art, but I don't see dependency parsing as a task in that suite of benchmarks, and NLP isn't my primary field, so I'm not up-to-date on the whether or not another method has outperformed it in the last three years. (I would be surprised if one hadn't.) At the time it was, and regardless, I don't doubt that the current state of the art similarly uses neural networks only as a subset of the complete method.

[1] https://www.aclweb.org/anthology/W04-0308

Edit: I can't seem to reply to the post below me. There isn't an entry CMU/Pittsburgh in your benchmark. The paper refers to the Stanford Dependency [SD] and Penn Chinese Treebank 5 [CTB5], and reports 90.9/85.7, respectively, for LAS scores. I'm not sure if any of the treebanks listed correspond to these, but these numbers do seem in line to at least be competitive regardless of which treebank is used.

> But I don't see dependency parsing as a task in that suite of benchmarks

That site is dedicated for dependency parsing task, scorecard represents results of various teams/models.

Obligatory:

Every time I fire a linguist, the performance of our speech recognition system goes up. -- Fred Jelinek

Dunno why you were downvoted, I heard this one multiple times during my classes and always had a good laugh :-)

Reply on news.ycombinator.com