It certainly helps, however, I think Aidan's suggestion goes into the direction of having an official dump distribution.

Imagine how many CO2 can be spared just by avoiding the computational resource to recreate this dump every time ones need it.

Besides, it standardise the dataset used for research purposes.

On Wed, Dec 18, 2019, 11:26 Marco Fossati fossati@spaziodati.eu wrote:

Hi everyone,

Benno (in CC) has recently announced this tool: https://tools.wmflabs.org/wdumps/

I haven't checked it out yet, but it sounds related to Aidan's inquiry. Hope this helps.

Cheers,

Marco

On 12/18/19 8:01 AM, Edgard Marx wrote:

...

+1

On Tue, Dec 17, 2019, 19:14 Aidan Hogan <aidhog@gmail.com mailto:aidhog@gmail.com> wrote:

Hey all,

As someone who likes to use Wikidata in their research, and likes to
give students projects relating to Wikidata, I am finding it more and
more difficult to (recommend to) work with recent versions of

Wikidata

...

due to the increasing dump sizes, where even the truthy version now
costs considerable time and machine resources to process and handle.

In

...

some cases we just grin and bear the costs, while in other cases we
apply an ad hoc sampling to be able to play around with the data and
try
things quickly.

More generally, I think the growing data volumes might inadvertently
scare people off taking the dumps and using them in their research.

One idea we had recently to reduce the data size for a student

project

...

while keeping the most notable parts of Wikidata was to only keep
claims
that involve an item linked to Wikipedia; in other words, if the
statement involves a Q item (in the "subject" or "object") not
linked to
Wikipedia, the statement is removed.

I wonder would it be possible for Wikidata to provide such a dump to
download (e.g., in RDF) for people who prefer to work with a more
concise sub-graph that still maintains the most "notable" parts?

While

...

of course one could compute this from the full-dump locally, making
such
a version available as a dump directly would save clients some
resources, potentially encourage more research using/on Wikidata, and
having such a version "rubber-stamped" by Wikidata would also help to
justify the use of such a dataset for research purposes.

... just an idea I thought I would float out there. Perhaps there is
another (better) way to define a concise dump.

Best,
Aidan

_______________________________________________
Wikidata mailing list
Wikidata@lists.wikimedia.org <mailto:Wikidata@lists.wikimedia.org>
https://lists.wikimedia.org/mailman/listinfo/wikidata

---

Wikidata mailing list Wikidata@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikidata

---

Wikidata mailing list Wikidata@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikidata

Hi all,

Yes, Benno's WDumper could be used for this purpose. The motivation for the whole project was very similar to what Aidan describes. We realised thought that there won't be a single good way to build smaller dump that would serve every conceivable use in research, which is why the UI let's users make custom dumps.

In general, we are happy to hear more ideas on how to build useful smaller dumps that would be interesting. We are also accepting pull requests.

Benno, could you add a feature to include only items with Wikipedia page (in some language, or in any language)?

Edgard, I don't think making this more "official" will be very important for most researchers. Benno spent quite some time on aligning the RDF export with the official dumps, so in practice, WDumper mostly produces a subset of the triples of the official dump (which one could also have extracted manually). If there are differences left between the formats, we will be happy to hear about them (a github issue would be the best way to report it).

As Benno already wrote, WDumper connects to Zenodo to ensure that exported datasets are archived in a permanent and citable fashion. This is very important for research. As far as I know, none of the existing dumps (official or not) guarantee long-term availability at the moment.

Cheers,

Markus

On 18/12/2019 13:37, Edgard Marx wrote:

It certainly helps, however, I think Aidan's suggestion goes into the direction of having an official dump distribution.

Imagine how many CO2 can be spared just by avoiding the computational resource to recreate this dump every time ones need it.

Besides, it standardise the dataset used for research purposes.

On Wed, Dec 18, 2019, 11:26 Marco Fossati <fossati@spaziodati.eu mailto:fossati@spaziodati.eu> wrote:

Hi everyone,

Benno (in CC) has recently announced this tool:
https://tools.wmflabs.org/wdumps/

I haven't checked it out yet, but it sounds related to Aidan's inquiry.
Hope this helps.

Cheers,

Marco

On 12/18/19 8:01 AM, Edgard Marx wrote:
 > +1
 >
 > On Tue, Dec 17, 2019, 19:14 Aidan Hogan <aidhog@gmail.com
<mailto:aidhog@gmail.com>
 > <mailto:aidhog@gmail.com <mailto:aidhog@gmail.com>>> wrote:
 >
 >     Hey all,
 >
 >     As someone who likes to use Wikidata in their research, and
likes to
 >     give students projects relating to Wikidata, I am finding it
more and
 >     more difficult to (recommend to) work with recent versions of
Wikidata
 >     due to the increasing dump sizes, where even the truthy
version now
 >     costs considerable time and machine resources to process and
handle. In
 >     some cases we just grin and bear the costs, while in other
cases we
 >     apply an ad hoc sampling to be able to play around with the
data and
 >     try
 >     things quickly.
 >
 >     More generally, I think the growing data volumes might
inadvertently
 >     scare people off taking the dumps and using them in their
research.
 >
 >     One idea we had recently to reduce the data size for a
student project
 >     while keeping the most notable parts of Wikidata was to only keep
 >     claims
 >     that involve an item linked to Wikipedia; in other words, if the
 >     statement involves a Q item (in the "subject" or "object") not
 >     linked to
 >     Wikipedia, the statement is removed.
 >
 >     I wonder would it be possible for Wikidata to provide such a
dump to
 >     download (e.g., in RDF) for people who prefer to work with a more
 >     concise sub-graph that still maintains the most "notable"
parts? While
 >     of course one could compute this from the full-dump locally,
making
 >     such
 >     a version available as a dump directly would save clients some
 >     resources, potentially encourage more research using/on
Wikidata, and
 >     having such a version "rubber-stamped" by Wikidata would also
help to
 >     justify the use of such a dataset for research purposes.
 >
 >     ... just an idea I thought I would float out there. Perhaps
there is
 >     another (better) way to define a concise dump.
 >
 >     Best,
 >     Aidan
 >
 >     _______________________________________________
 >     Wikidata mailing list
 > Wikidata@lists.wikimedia.org
<mailto:Wikidata@lists.wikimedia.org>
<mailto:Wikidata@lists.wikimedia.org
<mailto:Wikidata@lists.wikimedia.org>>
 > https://lists.wikimedia.org/mailman/listinfo/wikidata
 >
 >
 > _______________________________________________
 > Wikidata mailing list
 > Wikidata@lists.wikimedia.org <mailto:Wikidata@lists.wikimedia.org>
 > https://lists.wikimedia.org/mailman/listinfo/wikidata
 >

_______________________________________________
Wikidata mailing list
Wikidata@lists.wikimedia.org <mailto:Wikidata@lists.wikimedia.org>
https://lists.wikimedia.org/mailman/listinfo/wikidata

---

Wikidata mailing list Wikidata@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikidata

Hey all,

Just a general response to all the comments thus far.

- @Marco et al., regarding the WDumper by Benno, this is a very cool initiative! In fact I spotted it just *after* posting so I think this goes quite some ways towards addressing the general issue raised.

- @Markus, I partially disagree regarding the importance of rubber-stamping a "notable dump" on the Wikidata side. I would see it's value as being something like the "truthy dump", which I believe has been widely used in research for working with a concise sub-set of Wikidata. Perhaps a middle ground is for a sporadic "notable dump" to be generated by WDumper and published on Zenodo. This may be sufficient in terms of making the dump available and reusable for research purposes (or even better than the current dumps, given the permanence you mention). Also it would reduce costs on the Wikidata side (I don't think a notable dump would be necessary to generate on a weekly basis, for example).

- @Lydia, good point! I was thinking that filtering by wikilinks will just drop some more obscure nodes (like Q51366847 for example), but had not considered that there are some more general "concepts" that do not have a corresponding Wikipedia article. All the same, in a lot of the research we use Wikidata for, we are not particularly interested in one thing or another, but more interested in facilitating what other people are interested in. Examples would be query performance, finding paths, versioning, finding references, etc. But point taken! Maybe there is a way to identify "general entities" that do not have wikilinks, but do have a high degree or centrality, for example? Would a degree-based or centrality-based filter be possible in something like WDumper (perhaps it goes beyond the original purpose; certainly it does not seem trivial in terms of resources used)? Would it be a good idea?

In summary, I like the idea of using WDumper to sporadically generate -- and publish on Zenodo -- a "notable version" of Wikidata filtered by sitelinks (perhaps also allowing other high-degree or high-PageRank nodes to pass the filter). At least I know I would use such a dump.

Best, Aidan

On 2019-12-19 6:46, Lydia Pintscher wrote:

On Tue, Dec 17, 2019 at 7:16 PM Aidan Hogan aidhog@gmail.com wrote:

...

Hey all,

As someone who likes to use Wikidata in their research, and likes to give students projects relating to Wikidata, I am finding it more and more difficult to (recommend to) work with recent versions of Wikidata due to the increasing dump sizes, where even the truthy version now costs considerable time and machine resources to process and handle. In some cases we just grin and bear the costs, while in other cases we apply an ad hoc sampling to be able to play around with the data and try things quickly.

More generally, I think the growing data volumes might inadvertently scare people off taking the dumps and using them in their research.

One idea we had recently to reduce the data size for a student project while keeping the most notable parts of Wikidata was to only keep claims that involve an item linked to Wikipedia; in other words, if the statement involves a Q item (in the "subject" or "object") not linked to Wikipedia, the statement is removed.

I wonder would it be possible for Wikidata to provide such a dump to download (e.g., in RDF) for people who prefer to work with a more concise sub-graph that still maintains the most "notable" parts? While of course one could compute this from the full-dump locally, making such a version available as a dump directly would save clients some resources, potentially encourage more research using/on Wikidata, and having such a version "rubber-stamped" by Wikidata would also help to justify the use of such a dataset for research purposes.

... just an idea I thought I would float out there. Perhaps there is another (better) way to define a concise dump.

Best, Aidan

Hi Aiden,

That the dumps are becoming too big is an issue I've heard a number of times now. It's something we need to tackle. My biggest issue is deciding how to slice and dice it though in a way that works for many use cases. We have https://phabricator.wikimedia.org/T46581 to brainstorm about that and figure it out. Input from several people very welcome. I also added a link to Benno's tool there. As for the specific suggestion: I fear relying on the existence of sitelinks will kick out a lot of important things you would care about like professions so I'm not sure that's a good thing to offer officially for a larger audience.

Cheers Lydia

On Sat, 21 Dec 2019 at 17:25, Lydia Pintscher Lydia.Pintscher@wikimedia.de wrote:

On Thu, Dec 19, 2019 at 11:16 PM Aidan Hogan aidhog@gmail.com wrote:

...

• @Lydia, good point! I was thinking that filtering by wikilinks will

just drop some more obscure nodes (like Q51366847 for example), but had not considered that there are some more general "concepts" that do not have a corresponding Wikipedia article. All the same, in a lot of the research we use Wikidata for, we are not particularly interested in one thing or another, but more interested in facilitating what other people are interested in. Examples would be query performance, finding paths, versioning, finding references, etc. But point taken! Maybe there is a way to identify "general entities" that do not have wikilinks, but do have a high degree or centrality, for example? Would a degree-based or centrality-based filter be possible in something like WDumper (perhaps it goes beyond the original purpose; certainly it does not seem trivial in terms of resources used)? Would it be a good idea?

I think it's definitely worth exploring but I fear it needs someone to actually sit down and collect the different dumps use-cases and talk to people to figure out which part of the data they need. Based on that we could identify common patterns.

Yeah, there are a bunch of quite varied motivations for subsets. I have found the topic of Wikidata subsetting and data dumps coming up again and again. Most recently in a lifescience/bioinformations setting which is how we ended up collecting raw materials in the doc already shared here, https://docs.google.com/document/d/1MmrpEQ9O7xA6frNk6gceu_IbQrUiEYGI9vcQjDvT... but also in other domains. If people here care to drop use cases, thoughts and notes (*however scrappy*) into that doc I will make a pass over it to try to pull together a more readable summary of the various motivations for subsetting.

The work Adam wrote up at https://addshore.com/2019/10/your-own-wikidata-query-service-with-no-limits-... is also very relevant...

(I think this is something

that needs to be done but unfortunately can't dedicate time to it in the foreseeable future. https://phabricator.wikimedia.org/T46581 is a good place for people who want to help think it through.

That is also a fine place to record things! I don’t mean to fork the discussion. Maybe we could have a call for interested parties in the new year?

Dan

Cheers Lydia

-- Lydia Pintscher - http://about.me/lydia.pintscher Product Manager for Wikidata

Wikimedia Deutschland e.V. Tempelhofer Ufer 23-24 10963 Berlin www.wikimedia.de

Wikimedia Deutschland - Gesellschaft zur Förderung Freien Wissens e. V.

Eingetragen im Vereinsregister des Amtsgerichts Berlin-Charlottenburg unter der Nummer 23855 Nz. Als gemeinnützig anerkannt durch das Finanzamt für Körperschaften I Berlin, Steuernummer 27/029/42207.

---

Wikidata mailing list Wikidata@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikidata

Hi Aidan,

since DBpedia has been around for twelve years now, we spent the last 3 years intensively re-engineering to solve problems like this.

Last week, we finished the Virtuoso DBpedia Docker[1]  to work on Databus Collections[2],[3]. Databus contains different repartitions of all the datasets, i.e. Wikipedia/Wikidata extractions and external data. The idea here is that datasets or graphs are stored in a granular manner and then you make your own collection (DCAT Catalog) or re-use collections by others.

This will go in the direction to build 1 Billion derived Knowledge Graphs until 2025: https://databus.dbpedia.org/dbpedia/publication/strategy/2019.09.09/strategy...

We analysed a lot of problems in the GlobalFactSync project [5] and studied Wikidata intensively. Our conclusion here is that we will invert DBpedia by 180 degrees in the future. So instead of taking the main data from Wikipedia and Wikidata, we will take it from the sources directly, since all data in Wikipedia and Wikidata comes from somewhere else. So the new direction is LOD -> DBpedia -> Wikipedia/Wikidata via sameAs,equivalentClass/Property mappings.

This is not a solution for the dump size problem per se, because we are creating even bigger and more varied and domain-specific knowledge graphs and dumps via FlexiFusion. Besides the flexible source partitions, we offer a partition by property, where you can simply pick the properties you want for your knowledge graph and then docker-load it into the SPARQL store of choice. There is no manual yet, but this query gives you all 3.8 million birthdates of the new big fused graph: https://databus.dbpedia.org/yasgui/

PREFIX dataid: http://dataid.dbpedia.org/ns/core# PREFIX dataid-cv: http://dataid.dbpedia.org/ns/cv# PREFIX dct: http://purl.org/dc/terms/ PREFIX dcat:  http://www.w3.org/ns/dcat# SELECT DISTINCT ?file WHERE {      ?dataset dataid:version https://databus.dbpedia.org/vehnem/flexifusion/fusion/2019.11.15 .     ?dataset dcat:distribution ?distribution .     ?distribution http://dataid.dbpedia.org/ns/cv#tag 'birthDate'^^http://www.w3.org/2001/XMLSchema#string .     ?distribution dcat:downloadURL ?file . }

Where you can filter more here: https://databus.dbpedia.org/vehnem/flexifusion/fusion/2019.11.15

During the next year, we will include all European library data into the syncing process, several national statistical datasets and other data and refine the way to extract exactly the partition you need. It is an opportunistic extension to Linked Open Data, where you can select the partition you need independent of the IDs or vocab used.

-- Sebastian

[1] https://github.com/dbpedia/Dockerized-DBpedia

[2] https://forum.dbpedia.org/t/dbpedia-dataset-2019-08-30-pre-release/219

[3] https://github.com/dbpedia/minimal-download-client

[4] https://svn.aksw.org/papers/2019/ISWC_FlexiFusion/public.pdf

[5] https://meta.wikimedia.org/wiki/Grants:Project/DBpedia/GlobalFactSyncRE

On 19.12.19 23:15, Aidan Hogan wrote:

Hey all,

Just a general response to all the comments thus far.

• @Marco et al., regarding the WDumper by Benno, this is a very cool

initiative! In fact I spotted it just *after* posting so I think this goes quite some ways towards addressing the general issue raised.

• @Markus, I partially disagree regarding the importance of

rubber-stamping a "notable dump" on the Wikidata side. I would see it's value as being something like the "truthy dump", which I believe has been widely used in research for working with a concise sub-set of Wikidata. Perhaps a middle ground is for a sporadic "notable dump" to be generated by WDumper and published on Zenodo. This may be sufficient in terms of making the dump available and reusable for research purposes (or even better than the current dumps, given the permanence you mention). Also it would reduce costs on the Wikidata side (I don't think a notable dump would be necessary to generate on a weekly basis, for example).

• @Lydia, good point! I was thinking that filtering by wikilinks will

just drop some more obscure nodes (like Q51366847 for example), but had not considered that there are some more general "concepts" that do not have a corresponding Wikipedia article. All the same, in a lot of the research we use Wikidata for, we are not particularly interested in one thing or another, but more interested in facilitating what other people are interested in. Examples would be query performance, finding paths, versioning, finding references, etc. But point taken! Maybe there is a way to identify "general entities" that do not have wikilinks, but do have a high degree or centrality, for example? Would a degree-based or centrality-based filter be possible in something like WDumper (perhaps it goes beyond the original purpose; certainly it does not seem trivial in terms of resources used)? Would it be a good idea?

In summary, I like the idea of using WDumper to sporadically generate -- and publish on Zenodo -- a "notable version" of Wikidata filtered by sitelinks (perhaps also allowing other high-degree or high-PageRank nodes to pass the filter). At least I know I would use such a dump.

Best, Aidan

On 2019-12-19 6:46, Lydia Pintscher wrote:

...

On Tue, Dec 17, 2019 at 7:16 PM Aidan Hogan aidhog@gmail.com wrote:

...

Hey all,

As someone who likes to use Wikidata in their research, and likes to give students projects relating to Wikidata, I am finding it more and more difficult to (recommend to) work with recent versions of Wikidata due to the increasing dump sizes, where even the truthy version now costs considerable time and machine resources to process and handle. In some cases we just grin and bear the costs, while in other cases we apply an ad hoc sampling to be able to play around with the data and try things quickly.

More generally, I think the growing data volumes might inadvertently scare people off taking the dumps and using them in their research.

One idea we had recently to reduce the data size for a student project while keeping the most notable parts of Wikidata was to only keep claims that involve an item linked to Wikipedia; in other words, if the statement involves a Q item (in the "subject" or "object") not linked to Wikipedia, the statement is removed.

I wonder would it be possible for Wikidata to provide such a dump to download (e.g., in RDF) for people who prefer to work with a more concise sub-graph that still maintains the most "notable" parts? While of course one could compute this from the full-dump locally, making such a version available as a dump directly would save clients some resources, potentially encourage more research using/on Wikidata, and having such a version "rubber-stamped" by Wikidata would also help to justify the use of such a dataset for research purposes.

... just an idea I thought I would float out there. Perhaps there is another (better) way to define a concise dump.

Best, Aidan

Hi Aiden,

That the dumps are becoming too big is an issue I've heard a number of times now. It's something we need to tackle. My biggest issue is deciding how to slice and dice it though in a way that works for many use cases. We have https://phabricator.wikimedia.org/T46581 to brainstorm about that and figure it out. Input from several people very welcome. I also added a link to Benno's tool there. As for the specific suggestion: I fear relying on the existence of sitelinks will kick out a lot of important things you would care about like professions so I'm not sure that's a good thing to offer officially for a larger audience.

Cheers Lydia

---

Wikidata mailing list Wikidata@lists.wikimedia.org https://lists.wikimedia.org/mailman/listinfo/wikidata