The Discovery department's  work to improve search continues with a new
tool! We are asking volunteers to help choose, or discern, which search
results are the most relevant.
One way of improving search results relevance is to provide search results
from multiple search engines side-by-side for comparison. Participants pick
the best, most relevant results, which are then used to tune our own search
results. It's one way to help improve search with human assistance, by
showing articles that are most relevant to search queries. This system is
used by many R&D departments and gives great results.
Discernatron is a tool developed by the Discovery department for just this
sort of work. Visitors are asked to pick the most relevant results across
four different search result sets. The data is then used to help improve
our relevancy model for search. Screenshot at 
If you are interested in helping, you can access Discernatron at
https://discernatron.wmflabs.org/ and authenticate with your unified user
To learn more about the tool visit Mediawiki.org. 
Community Liaison - Discovery
We now have a weekly deployment window for WDQS each week on Monday
. This is great. I'd like to see how we can improve this a bit
I have been struggling to know if there is something planned for those
deployment windows, and what that content is. I can obviously blindly
deploy the latest commit each week and trust that everything is just
fine. That might actually be a good idea, but if that's the case, I
should work on completely automating it and remove myself from the
Other deployment window require the change initiator (the developer in
most cases) to add a description of the changes to be deployed to the
deployment page . I was actually expecting that and did not clarify
this point with Stas (my bad, I should not assume things).
@Stas (or others): would you be OK following this same process? Or
should we push automation a bit more and take myself out of the
Thanks for your idea!
Operations Engineer, Discovery
CCing the Search and Discovery list.
On Sun, May 8, 2016 at 12:24 PM, Stan Zonov <stanzon(a)gmail.com> wrote:
> I have been trying to gage the speed/efficiency of a database I have setup.
> In order to test it, I have filled it with a lot of wikipedia articles from
> a specific category (for example history). The database does multi-word
> queries and returns the articles that best match the multiword query. For
> example if I search up "history in Italy in the past 100 years" then the
> best matching articles should pop up.
> I was wondering if anyone has any advice how to form sample test queries to
> model realistic situations/queries. I don't think it would be fair to do
> random phrases (such as "banana the string") and wanted to model queries
> based on my data to test performance and correctness of output. Does anyone
> have any advice? How or Is this done at wikipedia?
> I have looked here
> but the data has been down for a while.
> Analytics mailing list
IRC (Freenode): HaeB
sad news, the completion suggester v2 was reverted from elastic 2.x
branch and is now delayed to elasticsearch v5.
Unfortunately I was planning to use important features in this version
to add realtime support.
I'll start to think about it and see if we can workaround and still
improve the current implementation.
I'm really sorry about that, I should have monitored the elasticsearch
repo more closely...
Stas and I just met to discuss how he can assist with search moving
forwards. Stas works on a lot of projects, so we wanted to guide and plan
his efforts together.
Given Stas's role as a technical liaison between the Wikidata and Search
teams, we started there. I mentioned how structured file metadata for
Commons is important for content discovery, given that it would potentially
enable us to do more with search on Commons by having access to structured
data. This was quite a popular item
on the community wishlist, where it was noted that the Wikidata Team would
be working on it. Discovery has some limited capacity to help with
architectural discussions, and Lydia and I have discussed helping in this
Stas is going to proceed with investigating T89733
<https://phabricator.wikimedia.org/T89733> "Allow ContentHandler to expose
structured data to the search engine", which is a huuuuge task which needs
a bit more definition to be actionable. Stas is continuing to work on the
Wikidata Query Service and also (for this quarter) supporting the
Performance Team in some of their efforts.
If you have any questions, let me know!
Lead Product Manager, Discovery
There seems be a data validation error in the Search schema. It's not super
pressing, but it should be fixed at some point. I created T134282
<https://phabricator.wikimedia.org/T134282> to track this.
---------- Forwarded message ----------
From: Marcel Ruiz Forns <mforns(a)wikimedia.org>
Date: 2 May 2016 at 04:48
Subject: Search schema validation errors
To: Dan Garry <dgarry(a)wikimedia.org>
We've observed that EventLogging's Search schema is receiving around 1400
events per hour that fail validation with the following error:
u'comp_suggest' is not one of ['prefix', 'fulltext',
It seems the client is sending a value that is not registered in the schema
for the comp_suggest field. 1400 ev/h is probably not a big share of all
Search events, but just a heads up in case you want to look into this.
You can easily analyze the error logs here:
*Marcel Ruiz Forns*
Lead Product Manager, Discovery
I've just finished my write-up for optimizing the languages that could
eventually be used for language detection on French Wikipedia. (Spanish,
Italian, and German are still to come.)
The full write-up
on corpus creation and clean up, performance stats, and more.
Briefly, about 15% of "low performing" queries (those with < 3 results) are
easily filtered junk, and 65% of the remainder are not in an identifiable
language (e.g., names, acronyms, more junk, etc.).
Based on a sample of 682 poor-performing queries on frwiki that are in some
language, about 70% are in French, 10-15% are in English, about 7-12% are
in Arabic, fewer than 3% are in Portuguese, German, and Spanish, and there
are a handful of other languages present.
Because of the relatively low percentage of low-performing queries that are
relevant, we will still need to do an A/B test before discussing deploying
this to frwiki. An A/B test on enwiki
<https://phabricator.wikimedia.org/T121542> in in the works at the moment.
The optimal settings for frwiki, based on these experiments, would be to
use the TextCat query-based models for French, English, Arabic, Russian,
Chinese, Armenian, Thai, Greek, Hebrew, Korean (fr, en, ar, ru, zh, th, el,
hy, he, ko), using the default 3000-ngram models.
Software Engineer, Discovery
This is a bad time for Wikidata Query Service. We had a few issues
last week to update WDQS and reload all data. During this data reload,
we are running on a single server. And now this server is starting to
I honestly do not know yet what is happening except that we see a lot
of file (pipes actually) left open and that WDQS stops responding.
Investigation is going on  and we'll let you know what we find.
Sorry for the inconvenience and thank you all for your patience!
Operations Engineer, Discovery