PanImages
A press release about an interesting project at the University of Washington:
A rose is a rózsa is a 薔薇: Image-search tool speaks hundreds of languages
PanImages is a tool that helps you to search for images across languages.
Searching for
“house” on Google images will get you different results than searching for tŷ or casa or maison.
PanImages takes all this to the next level, and automatically translates your search term into various languages, and then re-runs an ubersearch on Google Images and Flickr.
Pretty neat. But more interesting than the interface itself (to yours truly, anyway), is the work that’s going on behind the scenes to find the translated terms to broaden the query. The project is based on a paper:
Lexical Translation with Application to Image Search on the Web (pdf)
I just started taking a look at it, but there is some stuff in there about “translation graphs” that suggests all kinds of interesting applications.
UPDATE: I should add that I still find it compelling to note that there are more translations for “house” hidden in the interwiki links on the normal Wikipedia than there are to be found by parsing out the contents of Wiktionary: 45 versus 33, in this case.
OmegaWiki (a more marked up and semantic Wiktionary, may even replace it) lists 54 translations of house:
http://www.omegawiki.org/Expression:house
:)
Hi Robin,
Good point. Is the plan actually to replace Wiktionary with OmegaWiki? Whenever I look at the markup of Wiktionary it makes me cry.
To think that the folks who built PanImages had to parse Wiktionary stuff is pretty terrifying. I find the interface utterly bewildering. For instance, an example that is used extensively in the press release linked above, the Zulu word ifriji, which means “refrigerator,” isn’t even in the Zulu Wiktionary , as far as I can tell. But it’s in the Wiktionnaire.
OmegaWiki seems like a good idea, but I daresay it seems pretty complex to use. Sometimes I think what the world really needs, wikiwise, is nothing but a huge list of multilingual terms:
en: refrigerator
pt: geladeira
zu: ifriji
Of course, it depends on what the user wants; if they’re a good translator, a lot of times the dictionary entry is just a starting point anyway ― the process of choosing the right word then continues with web searches. I actually find myself mostly annoyed by things like etymologies when I’m trying to use a dictionary for translation.
Hi again,
Well said, that with the crying :). I don’t know what the politics are, but the OmegaWiki project (formerly WiktionaryZ) was started by some Wiktionary users to overcome the shortcomings of the current approach. It is built on top of the Wikidata project, which provides a more structured way to store data in Mediawiki. At the moment, OmegaWiki is still in alpha/experimenting stage. But I hope as soon as it gets stable, it will become the official Wiktionary. A bit of history can be found on OmegaWiki’s project page on Wikipedia.
I haven’t really used it much, but if it’s too complex at the moment, I’m sure it will be improved for the common use cases or be configurable (for example, only show me meanings, not etymology).
With the simple list approach, what do you do about homonyms? Just list multiple possible translations? In OmegaWiki, this is solved by having a “defined meaning” for each meaning and a definition alongside, so you can easily differentiate between homonyms. The simple approach is sure faster than the structured one – but why not do it right right from “the beginning” :)?
Hi Robin,
Thanks for your interesting thoughts.
Pretty much. What I personally would like to see is just lists. There is lots of research on automated word and phrase alignment of translations, and it outputs lists that are more or less like what you see below (but much messier). But as I say, even messiness doesn’t phase translators; anyone who knows American English well is going to recognize that there are tonal differences between “refrigerator” and “fridge,” for instance. Or if they don’t happen to be aware of the difference, they’ll be able to tease them out with some research.
The question of how to handle homonyms is hard. Come to think of it, is there some way to measure how common homonyms are in a given language? Do rates of homonymy vary across languages? (I would imagine that languages with very small phoneme inventories would probably have more homonyms than languages with large ones.)
It’s not possible to have a consistent, complete dictionary. Case in point:
Blue-green across cultures
But this problem is inescapable in any listing, even the listings that constitute OmegaWiki’s structure.
I guess what I have in mind could be articulated like this, to blockquote myself:
Which involves all kinds of handwaving and numerical questions undefined, but it seems to me that that kind of handwaving is different from saying “first we will set up a universal list of ideas, and the proceed to collect all translations of those ideas.” Perhaps I’m wrong, but that seems to be the approach OmegaWiki is taking.
Incidentally, the interwiki links between various Wikipedias (at least for nouns) still strike me as the closest, largest approximation to this ideal.