Comment by laGrenouille | Hacker News Reader

Comment by laGrenouille | Hacker News Reader

I generally agree with point made in the article that too much NLP research is focused only on English and a small number of other high-resourced languages. To me, this is part of a larger problem with natural language processing's obsession with "state-of-the-art" metrics and general abandonment of broader research in linguistics.

The situation for those who, like myself, work in applied linguistics is not actually so dire... At least when working with the languages that have a reasonable amount of training data. Decent enough treebanks exist for lemmatisation, POS-tagging, and dependency parsing for dozens of languages [0]. Fast tools such as spaCy (15 languages) [1] and udpipe (40+ languages) [2] are freely available and work well for most applied tasks. There are even decent word embeddings available trained on various versions of Wikipedia [3].

Of course, some of the issue is that these very tasks are biased towards the specific structures of Indo-European languages. However, for getting work done (building sentiment classifiers, document clustering, NER, ect.), currently available tools make it possible to work with a large proportion of the currently available data.

There is still a lot of work to be done own regards to computational research on non-English languages, but a lot of the problem right now is recognition of applied work by the top NLP conferences rather than a complete lack of quality work currently being done.

[0] https://universaldependencies.org

[1] https://spacy.io/models

[2] https://ufal.mff.cuni.cz/udpipe/models

[3] https://github.com/facebookresearch/fastText/blob/master/doc...

One thing i think about is what where would NLP and linguistics research gone if another language (other than english) was the predominant language. Granted there is quite of bit interesting fundamental research in other countries and languages. But it seems many of problems and approaches developed in language understanding and NLP target specific idiosyncracies of the English language. If NLP researched with say German (which has a deep and overly explicit vocabulary) or Hindi or Mandarian, would the fundamental NLP approaches and problem areas be different and potentially better?

I recommend you to check out NLP research regarding the Czech language. It has high quality researchers keen on making use of the language's features (it's a highly regular and expressive grammar). Sadly not much research is done, but what is done is interesting.

Agreed. The way I see it, current research directions have focused on approaches that work best with large datasets because such datasets exist for English, and because English just happens to be a language that is possible to represent well by training on large datasets.

It helps of course that English is the de facto lingua franca of most of the technologically developed world (if not the rest of the world also). If another language was the lingua franca, current approaches would not work so well. The article makes a point about how we assume that pre-trained embeddings encode all relevant information about a language when this may not be the case for arbitrary languages. The article also makes a point about the morphological simplicity of English which is of course a great asset for a lingua franca, and also simplifies the language modelling task. Imagine if the lingua franca was, e.g. ancient Greek (as it was, in ancient times, all around the Mediterrannean) or something with similar morphological intricacies (and Greek is still an Indo-European language; I just don't know much about, e.g. Finno-Ugric to give an example about them).

Is it practical to do NLP research using a language you're not fluent in? Like, does most research require the experimenter to be able to judge fluency and ambiguity themselves?

For sure. Many NLP scientific benchmarks are multilingual today and you need to evaluate on X languages. You are of course not expected to be fluent in all of them. Most of the deep learning based approaches are quite language-universal, i.e. you have a pre-trained language model or word embeddings and you just slap a sequence modeling model on top of that and you let the neural network do the magic.

However, you can still incorporate language-specific knowledge if you want to get the best performance for that one particular language. This is mainly true for academic research. If you are developing a rule-based or keyword-based application (which is still often the case outside of academia), it can actually help to be fluent to get the rules right.

Reply on news.ycombinator.com