Using formal grammar for translation

[voron]

Are there any freely available implementations on the net?

I was thinking about implementing an automatic translation service from English to Turkish this way:
1) use a grammar parser for English (such as Link grammar parser, http://www.link.cs.cmu.edu/link/submit-sentence-4.html)
2) implement a converter that converts the syntax tree from English into Turkish
3) provide translations for individual words using a online dictionary such as tureng.com (they even mark parts of speech), and generate endings using morhological information. I realize there is going to be ambuiguity, but I expect that even selecting the first word for translation would produce decent results.

Apparently this has already been done in research projects from English to Turkish, Persian and Indonesian, as this wikipedia article explains:
https://en.wikipedia.org/wiki/Link_grammar#Applications

I also found a few research papers in PDF such as this:
http://www.cs.bilkent.edu.tr/~ilyas/PDF ... thesis.pdf

Why aren't these works seeing any real life applications? Google translator from English to Turkish totally sucks, and I am inclined to think that the translation approach using the syntax analysis is going to be at least somewhat better (if not significantly better).

[vijayjohn]

Because link grammar parsers also suck. I actually tried doing something like this myself a long time ago. It doesn't work.

[voron]

The PDF that I gave the link to says they have 82% of sentences parsed correctly for Turkish. It looks like a high percent.

I'd still like to see how much it sucks. Do you know if there is a free demo anywhere? (between any pair of languages)

[vijayjohn]

:(
The PDF that I gave the link to says they have 82% of sentences parsed correctly for Turkish. It looks like a high percent.

It isn't. It's basically the same sorts of results that every paper on natural language processing reports. Sure, a lot of them report lower percentages, too, but it's not significantly higher AFAICT than something like Google Translate.

I'd still like to see how much it sucks. Do you know if there is a free demo anywhere? (between any pair of languages)

Nope, sorry.

Bear in mind, of course, that I'm also generally pretty cynical about stuff like this!

[linguoboy]

Linguists have only been trying this approach for, what, the last fifty years or so? If they could've gotten it to work well, they would've already. You just can't divorce syntax from semantics when attempting a translation; that was amply demonstrated back in the 70s. It's the reason I had given up on machine translation until the statistical approach emerged.

If you believe otherwise, I suggest you conduct an experiment of your own: Do what you suggest and manually translate a few sentences using the approach of simply taking the first dictionary entry for each lexeme and see what you end up with. For good measure, run the same sentences through Google Translate and compare them.

[voron]

Were there attempts to combine the two approaches? Google Translate can guess the right translation if there is a word with several meanings, but it messes the syntax so badly that the sentence becomes unreadable. What if we translated lexemes with the statistical approach, and dealt with the word order using grammar parsers?

[vijayjohn]

Here's the thing: Link grammars are not developed by people who actually know linguistics, nor are any automatic parsers I can recall. They're developed by programmers, and FWIR the parsers that exist are not effective.

Most linguists don't know how to program, very few programmers know any linguistics, and no one who knows both actually uses linguistics to create better technology (and if you try doing it yourself, they're not interested in collaborating with you). Instead, they're just endlessly fascinated by the technology, don't fully realize what the problems with it are, and ignore or dismiss the ones they know of. Collaboration between linguistics and programmers has been attempted in the past (several decades ago IIRC) but has never ended well; there's an old, nasty joke among such programmers that performance improved once they fired the linguist. Last I checked or heard, people who work in the area of NLP were also slow to realize that the problems they are trying to deal with are fundamentally linguistic in nature.

[voron]

It's a sad and a bit surprising since the area of NLP can potentially bring you $$$, and research centers of top IT companies can afford spending lots of money of finding top specialists and giving them good motivation (read: salary).

For example, there is a department of NLP at Google's research center:
https://research.google.com/pubs/Natura ... ssing.html

Maybe it's just a question of time until there is a generation of specialists who are good in both linguistics and programming.

[vijayjohn]

It's a sad and a bit surprising since the area of NLP can potentially bring you $$$

It can, but it requires you to do the dumb stuff they're currently doing. Any serious attempt to integrate linguistics into it would be a revolution that that field is just not ready for yet.

Maybe it's just a question of time until there is a generation of specialists who are good in both linguistics and programming.

Hopefully. I feel it's taking way too much time for (certain) people to notice that the study of linguistics potentially has practical value. There's no question that programming gets you a job.

[linguoboy]

Interestingly, I once got the offer of a programming job solely the basis of my linguistics degree.

[vijayjohn]

I think I've heard of that happening before, actually. I suppose there are various possible reasons for that (although perhaps my dad and/or brother would know better).