Property suggester data treats P279 as classifying, while the PropertySuggester extension does not
Closed, ResolvedPublic
Actions

Assigned To

Authored By

	daniel
	Jun 20 2017, 5:28 PM

Description

PropertySuggester on wikidata.org does not treat P279 as classifying, while the suggester data is generated under the assumption that P279 is classifying. This inconsistency may lead to sub-optimal suggestions on Items that use P279.

Evidence

PropertySuggester's extension.json sets PropertySuggesterClassifyingPropertyIds to [ 31 ]: https://github.com/Wikidata-lib/PropertySuggester/blob/master/extension.json. Configuration for the live site does not override this, see https://phabricator.wikimedia.org/source/mediawiki-config/browse/master/wmf-config/Wikibase-production.php (it only overrides it for test.wikidata.org). Or we may test this on the live site first, before changing the defaults.

However, PropertySuggester-Python's analyzer.ini sets it to 31,279, see https://github.com/Wikidata-lib/PropertySuggester-Python/blob/master/propertysuggester/analyzer/analyzer.ini. A look at the wbs_propertypairs table confirms that this configuration is used to produce the data for the live site.

Proposal

Add 279 to PropertySuggesterClassifyingPropertyIds in PropertySuggester's extension.json. This makes the default config consistent, and it causes the life site to treat P279 as classifying, allowing it to use the data in wbs_propertypairs correctly. Settings for labs and test should be adjusted accordingly. We may want to test this on the live site first, before changing the defaults, though.

We could also do it the other way around by removing 279 from analyzer.ini, but that would require the data to be re-generated. Also, treating P279 (subclass of) as classifying seems sensible.

Related Objects
Search...

Status	Assigned	Task
Resolved	hoo	T168434 Property suggester data treats P279 as classifying, while the PropertySuggester extension does not
Resolved	daniel	T169060 Set $wgPropertySuggesterClassifyingPropertyIds to [ 31, 279 ] for wikidata.org
Resolved	daniel	T169057 Change default value of $wgPropertySuggesterClassifyingPropertyIds to [ 31, 279 ]
Resolved	daniel	T169058 Temporarily set $wgPropertySuggesterClassifyingPropertyIds to [ 31 ] for wikidata.org

Event Timeline

daniel created this task.Jun 20 2017, 5:28 PM

Restricted Application added a subscriber: Aklapper. · View Herald TranscriptJun 20 2017, 5:28 PM

daniel triaged this task as Medium priority.Jun 20 2017, 5:29 PM

Sjoerddebruin awarded a token.Jun 20 2017, 5:30 PM

Let's add it and see if the suggestions improve.

daniel added a project: Wikidata-Former-Sprint-Board.Jun 21 2017, 9:07 AM

daniel updated the task description. (Show Details)

@Lydia_Pintscher I can make a config patch, but the config change should be announced in advanced, and we should ask people to test and comment. We should also make sure we can get it reverted quickly if need be.

I would love to help testing this. :)

I think we can consider this a bugfix. Why do you think it needs announcement in advance?

In T168434#3367629, @thiemowmde wrote:

I think we can consider this a bugfix. Why do you think it needs announcement in advance?

We got a few messages in the past of people who said "this used to show up in the suggestions and now it doesn't anymore". We need to check if the impact is not too much, but I highly doubt it.

daniel mentioned this in T169057: Change default value of $wgPropertySuggesterClassifyingPropertyIds to [ 31, 279 ].Jun 28 2017, 11:43 AM

daniel created subtask T169057: Change default value of $wgPropertySuggesterClassifyingPropertyIds to [ 31, 279 ].

daniel mentioned this in T169060: Set $wgPropertySuggesterClassifyingPropertyIds to [ 31, 279 ] for wikidata.org.Jun 28 2017, 11:49 AM

daniel created subtask T169060: Set $wgPropertySuggesterClassifyingPropertyIds to [ 31, 279 ] for wikidata.org.

daniel removed a subtask: T169057: Change default value of $wgPropertySuggesterClassifyingPropertyIds to [ 31, 279 ].

Is the below related ?

It seems that P31/P279 gets suggested on items that only have coordinates (P625), but not on items that only have P641 (sport). I'd expect it to appear on any item that lacks it.

In T168434#3430396, @Esc3300 wrote:

Is the below related ?

It seems that P31/P279 gets suggested on items that only have coordinates (P625), but not on items that only have P641 (sport). I'd expect it to appear on any item that lacks it.

Sport is removed from the propertypairs datafile because it is used on various types of items, while coordinates is only used on geographical items.

That makes sense for additional suggestions, but it still doesn't explain why P31/P279 aren't offered at all.

Try adding a statement to any item on these reports: https://www.wikidata.org/wiki/Category:P641_only_reports

In T168434#3430460, @Esc3300 wrote:

That makes sense for additional suggestions, but it still doesn't explain why P31/P279 aren't offered at all.

Try adding a statement to any item on these reports: https://www.wikidata.org/wiki/Category:P641_only_reports

It doesn't have something to base suggestions on. My opinion is that it should fallback to the initial list (that could use some additions imo, like image), but that is unrelated to this task.

Ok. I made T170422

daniel closed subtask T169060: Set $wgPropertySuggesterClassifyingPropertyIds to [ 31, 279 ] for wikidata.org as Resolved.Jul 31 2017, 3:25 PM

Restricted Application added a subscriber: PokestarFan. · View Herald TranscriptJul 31 2017, 3:25 PM

This seems resolved now, impact not that big.

Property suggester data treats P279 as classifying, while the PropertySuggester extension does notClosed, ResolvedPublicActions