Hi All,
Brent Hecht here :-) This has been a really interesting discussion, and I wanted to chime
in with a few notes.
The 99.2% is based on a quick script I wrote that looked at reciprocity among a sample of
interlanguage links (ILLs) in 25 languages to address some questions that Denny and I
brainstormed. It did not consider commented links, redirects, etc. However, in my lab’s
published work that takes much more involved approaches to this problem, we also find that
complex interlanguage link situations are the obvious minority, with simple 1:1
relationships being the norm. For instance, only 1% of connected components of the ILL
graph (groups of articles linked together by ILLs) had more than one article per language
edition in our 25-language dataset.
There are some important details to note here. For instance, most connected components
contain only one article (most concepts are covered by only a single language edition) and
non-1:1 cases by definition involve more articles than 1:1 cases (all other things being
equal). Also, general concepts of global interest are likely disproportionately
represented in the non-1:1 situations (e.g. river, canal, high school, diplomacy).
That said, given our data, I also think Denny is spot on with the “let’s start with the
1:1s” approach to building Wikidata, with solutions for more complex situations coming
later. These solutions could be fascinating and important, but make sense as a second
step, IMHO. Given that each language edition will be able to pick and choose from
statements (last I checked, at least), this might provide additional flexibility as well,
allowing greater variation to be included in the 1:1 model.
If interested, I'd encourage folks to check out our CHI 2012 paper [1], as well as
some excellent work done by Gerard de Melo and Gerhard Weikum that preceded us [3]. De
Melo and Weikum establish an interesting taxonomy for causes of non-1:1 links: conceptual
drift, different granularities, and mistakes made by editors.
In my view, perhaps a greater problem is the one of missing interlanguage links, which I
hope Wikidata’s popularity will help to solve. We’ve done some work to show that missing
links can be somewhat substantial between certain language editions [2], although that was
based on data from 2009.
It's important to note, too, that some of the differences in coverage of a given
concept across articles in different languages is addressed not with ILLs, but simply by
describing concepts differently in each language edition. We call this "sub-concept
diversity", and it can be substantial [2]. Our CHI 2012 paper describes a system we
built, Omnipedia, that allows folks to browse the content about a single concept in 25
language editions. We’re hoping to launch the system sometime soon, but we have some
practical considerations to deal with first (funding, finishing my thesis, etc. :-)).
Lastly, I've been digging into the social science of this stuff a bit lately and many
folks believe that, as Ziko said, different languages "divide knowledge in different
ways" (even apart from any effects introduced in the Wikipedia context specifically).
For instance, the linguist Anna Wierzbicka talks about the granularity differences as
"cultural elaboration" and has all sorts of fun examples in her book
"Understanding Cultures Through Their Keywords". You can also make arguments
about this from a geographic perspective (my social science roots), psycholinguistics, and
I'm sure other fields as well. This stuff perhaps explains some of the 1% of non-1:1
concepts, as well as some of the sub-concept diversity, although I am still
brainstorming.
In any case, hopefully this helps some! Happy to answer questions. Thanks again for a
great discussion.
- Brent
p.s. Don't forget to support Denny and crew in the Knight News Challenge proposal :-)
:
http://newschallenge.tumblr.com/post/25575917516/wikidata-as-a-central-free…
Brent Hecht
Ph.D. Candidate in Computer Science @ Northwestern University
Asst. Prof. of Comp. Sci @ Univ. Minnesota beginning 2013
w:
http://www.brenthecht.com
e: brent(a)u.northwestern.edu
t: @bhecht
[1] Bao, P., Hecht, B., Carton, S., Quaderi, M., Horn, M. and Gergle, D. 2012.
Omnipedia: Bridging the Wikipedia Language Gap. CHI ’12: 30th International Conference on
Human Factors in Computing Systems (2012).
[2] Hecht, B. and Gergle, D. 2010. The Tower of Babel Meets Web 2.0: User-Generated
Content and Its Applications in a Multilingual Context. CHI ’10: 28th International
Conference on Human Factors in Computing Systems (Atlanta, GA, 2010), 291–300.
[3] de Melo, G. and Weikum, G. 2010. Untangling the Cross-Lingual Link Structure of
Wikipedia. ACL ’10: 48th Annual Meeting of the Association for Computational Linguistics
(Uppsala, Sweden, 2010).
On Jun 26, 2012, at 7:56 AM, Denny Vrandečić wrote:
I got the number from Brent Hecht, a researcher at
Northwestern, who
has a number of great papers published on Wikipedia-related topics.
CC-ing him, so he knows I am blam.., er, referencing him :)
Cheers,
Denny
2012/6/26 Martijn Hoekstra <martijnhoekstra(a)gmail.com>om>:
This number, 99.2% was also mentioned on the
Berlin Hackathon. It
sounds much higher than what my (very scientifically relevant,
obviously) gut feeling tells me. Could you indicate where this number
is coming from?
On Tue, Jun 26, 2012 at 2:45 PM, Denny Vrandečić
<denny.vrandecic(a)wikimedia.de> wrote:
Ziko,
it does not jeopardize the Wikidata goal -- the current language link
system won't be switched off, but can be further used. Everything that
is working currently will still be possible afterwards. Wikidata can
still be used to represent the 99.2% of language links that are simple
-- this would still be a huge improvement over the current state.
As soon as these are out of the way, we can think about if and how to
extend the system in order to deal with the rest.
Cheers,
Denny
2012/6/25 Ziko van Dijk <vandijk(a)wmnederland.nl>nl>:
Hello,
So may I guess that "double links" are usually the result of a
Wikipedian who was not sure which language link to set, so in doubt,
he simply put in the language links for two different articles?
And in general, is it imagineable that different languages divide the
knowledge in different ways, which could jeopardize the whole goal of
Wikidata unifiying the language links?
Kind regards
Ziko
2012/6/25 Delirium <delirium(a)hackish.org>rg>:
> Thanks for this list. For the languages I know, I've started going through
> and fixing ones that are clearly wrong. If a number of people do that, that
> should improve the general quality/consistency of interwiki links. I second
> the other comment that it'd be nice if the parsing could be re-run to
> exclude commented-out links, but the list is still useful as is.
>
> There are some difficult cases, though, when languages make different
> choices on how to group subjects, so the articles aren't actually in 1-to-1
> correspondence. For example, the English article [[en: Móði and Magni]]
> unsurprisingly has two outgoing interwiki links, when linking to languages
> that split them, such as [[da:Magni]] and [[da:Modi]]. It's not clear what
> to do about these cases.
>
> Best,
> Mark
>
>
> On 6/25/12 12:29 PM, Denny Vrandečić wrote:
>>
>> Hi all,
>>
>> I ran some analysis last week, to get some numbers out of the
>> Wikipedia language links. One type of reports that were generated was
>> the list of all articles in the main namespaces of the Wikipedias that
>> link to more than one article in another language edition of Wikipedia
>> (so called double language links). There are not that many of them
>> (about 19,000 in total), split by language, all available here:
>>
>> <http://simia.net/languagelinks/>
>>
>> Double language links are not errors per se, but they contain a few
>> nuisances
>> * they lead to two links in the language links list that just look the
>> same (you have to hover over them to see that they link to different
>> languages), which is not really optimal from the user experience side
>> * they are not saved in the langlinks table and thus are ignored in
>> certain reports and also in the respective export
>>
>> I am not sure how to reach out to the respective Wikipedia
>> communities, or if I should at all. Should I post to their respective
>> version of the village pump? Remembering from the time I was active on
>> the Croatian Wikipedia, I would have appreciated that list to check
>> the entries. I reckoned the wikipedia-l list would be the right place,
>> but that list looks rather dead.
>>
>> Cheers,
>> Denny
>>
>
>
> _______________________________________________
> Wikimedia-l mailing list
> Wikimedia-l(a)lists.wikimedia.org
> Unsubscribe:
https://lists.wikimedia.org/mailman/listinfo/wikimedia-l
--
-----------------------------------------------------------
Vereniging Wikimedia Nederland
dr. Ziko van Dijk, voorzitter
http://wmnederland.nl/
Wikimedia Nederland
Postbus 167
3500 AD Utrecht
-----------------------------------------------------------
_______________________________________________
Wikimedia-l mailing list
Wikimedia-l(a)lists.wikimedia.org
Unsubscribe:
https://lists.wikimedia.org/mailman/listinfo/wikimedia-l
--
Project director Wikidata
Wikimedia Deutschland e.V. | Obentrautstr. 72 | 10963 Berlin
Tel. +49-30-219 158 26-0 |
http://wikimedia.de
Wikimedia Deutschland - Gesellschaft zur Förderung Freien Wissens e.V.
Eingetragen im Vereinsregister des Amtsgerichts Berlin-Charlottenburg
unter der Nummer 23855 B. Als gemeinnützig anerkannt durch das
Finanzamt für Körperschaften I Berlin, Steuernummer 27/681/51985.
_______________________________________________
Wikimedia-l mailing list
Wikimedia-l(a)lists.wikimedia.org
Unsubscribe:
https://lists.wikimedia.org/mailman/listinfo/wikimedia-l
_______________________________________________
Wikimedia-l mailing list
Wikimedia-l(a)lists.wikimedia.org
Unsubscribe:
https://lists.wikimedia.org/mailman/listinfo/wikimedia-l
--
Project director Wikidata
Wikimedia Deutschland e.V. | Obentrautstr. 72 | 10963 Berlin
Tel. +49-30-219 158 26-0 |
http://wikimedia.de
Wikimedia Deutschland - Gesellschaft zur Förderung Freien Wissens e.V.
Eingetragen im Vereinsregister des Amtsgerichts Berlin-Charlottenburg
unter der Nummer 23855 B. Als gemeinnützig anerkannt durch das
Finanzamt für Körperschaften I Berlin, Steuernummer 27/681/51985.