On Sat, Aug 15, 2015 at 1:17 PM Lydia Pintscher <
lydia.pintscher(a)wikimedia.de> wrote:
On Aug 15, 2015 14:06, "Magnus Manske" <magnusmanske(a)googlemail.com>
wrote:
On Sat, Aug 15, 2015 at 7:38 AM Lydia Pintscher <
lydia.pintscher(a)wikimedia.de> wrote:
>
> On Sat, Aug 15, 2015 at 3:43 AM, Dan Garry <dgarry(a)wikimedia.org>
wrote:
> > I've seen arguments on both sides
here. Some say automatically
generated
> > descriptions are not good enough. Some
say they are. Why don't we
gather
some data on this and use that to decide what's
right? :-)
Please do. Especially pay attention to languages other than English
though. Because even if we get algorithms to write good descriptions
for English are we going to do the same for all the other languages?
Especially those where grammar is tricky and Wikidata doesn't even
have the necessary information to make the grammar right? The other
tricky side is determining why something is actually notable. That's
not a trivial thing to determine based on the data we have.
And you know very well that (AFAIK) I am the only one who actually
worked on this,
in a tiny fraction of my spare time, and I only speak
German and English.
The /real/ questions here are:
1. The language that are actually implemented, are they returning
descriptions
that are good/OK/bad/plain wrong
2. What could be achieved, on the existing or
similar infrastructure, in
a short period of time, if we drive to get code snippets
(or equivalent)
for other languages from volunteers?
3. What could be achieved, medium/long term, if
we had a proper linguist
to work on the problem? Or someone who has worked with
multi-language text
generation before?
I've just been winging it so far. Current auto-descriptions are not the
best
we can do. They are, frankly, the WORST we can do. This is a starting
point, not the end product.
Yeah I understand. And this is not a criticism of your work. I think it is
actually rather cool. It is questioning if it is a good idea to continue to
push it to get into production on Wikipedia on a large scale.
With that, I agree wholeheartedly.
There might be a point of doing an "extended prototype" though, before
going to production (as much as I'd like that). What languages would be
easy, hard, impossible? Would this work as a stand-alone project (e.g.
dedicated VM), or as an extension of wikibase (flexibility vs. convenient
integration)? What open source code is already out there we could use?
Anyone in WMF/chapters who has experience in text generation? Anyone in
WMF/chapters who speaks a "small" language who could help set up an example
generator for that? What are the major item "classes" on Wikidata to be
covered with special code, beyond the obvious "human bio"?
And we'd need someone to run this. As much as I'd like to, I'm stretched
too thin as it is...