Wikidata:Property proposal/agent of action

From Wikidata
Jump to navigation Jump to search

agent of action

[edit]

Originally proposed at Wikidata:Property proposal/Generic

   Withdrawn
Descriptionthing that does the action
Data typeItem
Domainaction (Q4026292)
Example 1German December 16 suprise attack in the Battle of the Bulge (Q116504918)agent of actionArmy Group B (Q157572) Source
Example 2Opening of Tokyo 2020 games (Q116504974)agent of actionNaruhito (Q217096) Source
Example 3Johann Philipp Reis demonstration of the Reis telephone to the Physical Society of Frankfurt (Q116504999)agent of actionJohann Philipp Reis (Q77124) Source
Example 4crime (Q83267)agent of actioncriminal (Q2159907)
Example 5competition (Q841654)agent of actioncontestant (Q5165152)
Example 6telephone call (Q2296401)agent of actioncaller (Q113293705)
See alsohttps://schema.org/agent

Motivation

[edit]

I would like to create a data model to describe notable actions agents have made that are described in various Wikimedia articles. We should allow users to document actions so that they can be used to create timelines of events that can then be easily translated. They can also be used as a source to generate detailed Wikipedia article content for Abstract Wikipedia.

This property is the first to be proposed of the data model and follows the Schema.org data model for actions: https://schema.org/Action

participant (P710) exists, however that's usually used usually for events and not actions. It also requires that you use object of statement has role (P3831) to specify the role of the participant. For a relationship as critical and common as an agent is to the action they perform, we should have a dedicated property and not be required to add object of statement has role (P3831)agent (Q24229398) to every single agent statement. Lectrician1 (talk) 22:08, 30 January 2023 (UTC)[reply]

Discussion

[edit]

┌────────────────────────────────────────────────────────────────────────────────────────────────────┘ The rationale for splitting object of action and object class of action was expressed by User:BlaueBlüte on this earlier version of that proposal. Technically, they are right that without an instance–class distinction, some statements could be ambiguous to a non-human; however, that is also true of lots of existing properties, e.g. participant (P710), physically interacts with (P129), location (P276), and probably a hundred others. I would say there's no particularly good reason for such a property split on this proposal, but I support whatever gets it over the line, which the approach I took on the "object [class] of" proposal. Swpb (talk) 16:43, 6 August 2024 (UTC)[reply]

1) I understand that your objection is the mixing of instances and classes in the subject and object slots of the property. What I don't see is what problem you think that causes. As I pointed out, dozens of properties already function that way without issue. A problem would be the sort of ambiguity where it is unclear if a particular subject and/or object should be interpreted as instance or class. I am asking you for a single example of such a statement.
2) Fine-grained roles may be distinguished with object of statement has role (P3831) as I described above: Where a type of action permits multiple agent roles (e.g. an auction in which both the lot owner and the auction house are agents, but with different fine-grained roles), object of statement has role (P3831) can specify the role (just as it may now with participant (P710), but with the benefit that, unlike participant (P710), an unqualified "agent of action" statement does not permit its object to be an object of the action). I imagine you would rather see a multitude of properties like "owner of auction lot" and "auction house" that would enable that information to be expressed without the use of a qualifier. Maybe there is a case to be made for such properties, but I don't think you appreciate the context in which we are working: right now, information about agency is often only expressible with great ambiguity, especially when the action being described is the value of a main statement, with its agent identified by a qualifier. At the moment, participant (P710) and of (P642) are the only way to do this; the former only applies to people/organizations and, as I noted, is more ambiguous than "agent of action" with respect to role, and the latter has about a thousand meanings and is essentially useless for any sort of automated interpretation. Introducing "agent of action" greatly improves the expressibility of this information, while doing nothing to impede the introduction of more specific sub-properties later. In fact, having "agent of action" widely deployed will make it much easier to query for statements that may be candidates for migration to any future, fine-grained agent-role properties.
3) Yes, defender (Q111729140), victim (Q1851760), and military target (Q6518482) are roles, but they are also classes of objects that fill those roles. You won't see a general statement Army Group B (Q157572)instance of (P31)attacker (Q31924059), because that item is not always an attacker, but in the context of the statement about the 12/16 attack, it certainly is an instance of the class "attacker". The "rule" statement means that items in that position can be assumed to take on that role in context, not that they must always belong to that class. If that context-dependent class membership is a problem, we can very easily revise the description text of object class of action (P12913) to explicitly allow it: how about "class of objects (including substances) to which an action may occur, or role that the object of an action will fill in the context of that action"? Swpb (talk) 15:12, 13 August 2024 (UTC)[reply]
There is as well a problem with the domain. The examples are not just actions, but also classes of actions. This needs to be addressed. And the comments in a concern with the related property proposal “object of action” are also germane here, indicating that two properties are required.
This is not to say that the proposal cannot be salvaged, but work needs to be done on it. Peter F. Patel-Schneider (talk) 20:27, 13 August 2024 (UTC)[reply]
We could potentially use applies to use with property (P11527) as a qualifier to clear this up? I.e. crime (Q83267)agent of actioncriminal (Q2159907)applies to use with property (P11527)occupation (P106) -wd-Ryan (Talk/Edits) 02:26, 14 August 2024 (UTC)[reply]
@Wd-Ryan I don't think so. Consider Richard Nixon (Q9588) again. This item is very famous so the information there should be relatively complete. But there is no criminal (Q2159907) given for occupation (P106). The issue is that participation in an event very seldom affects the information stored on its object or agent because the participation is generally not sufficiently important to be noted there. Consider military offensive (Q2001676). The role of "agent (class) of action" there would be attacker (Q31924059). But the values of "agent of action" on instances of military offensive (Q2001676), such as Petsamo–Kirkenes Offensive (Q705222) with "agent of action" Soviet Union (Q15180), would not usually be instances of attacker (Q31924059) or even related to attacker (Q31924059). Peter F. Patel-Schneider (talk) 09:59, 14 August 2024 (UTC)[reply]
Politics aside, if Richard Nixon (Q9588) committed a crime, he would definitionally be a "criminal". This seems more like an error in the item's model. I don't think the value for Petsamo–Kirkenes Offensive (Q705222) would be Soviet Union (Q15180), but instead a specific organization or military unit that does attacking. -wd-Ryan (Talk/Edits) 16:16, 14 August 2024 (UTC)[reply]
Peter, I addressed your first concern in my point #3, and your second concern in point #1, of my response to Andrea above; I'd appreciate if you could engage with those responses and point out where you find them insufficient. Thanks, Swpb (talk) 15:07, 14 August 2024 (UTC)[reply]
I do not find your point #3 relevant at all. Sure, things like criminal (Q2159907) may be classes as well as occupations, but there is no hint in Wikidata that Richard Nixon (Q9588) is related to criminal (Q2159907) in any way. How then, is one to determine whether Richard Nixon (Q9588) is the kind of thing that can be the agent of a crime (Q83267)? I stated this concern in my original oppose message.
Your point #1 asks for examples where there can be a mixup between instances and the class. This problem was already pointed out for object of action and the result was that two properties were created, object of action (P12912) that provides the actual value for an action or instances of an action class and object class of action (P12913) that provides a class from which the value is taken. So window cleaning (Q3124765) uses object class of action (P12913) but circumnavigation of Earth (Q3533809) uses object of action (P12912), eliminating a nasty kind of ambiguity. There are indeed action instances that act on classes, for example the design of Volkswagen Passat B2 (Q30600) or statistical studies of criminal (Q2159907). There are also action instances whose agent is a class, for example any activity performed by a group of humans (Q16334295).
The two-property solution for objects of actions does overload both object of action (P12912) and object class of action (P12913), which is not generally a good idea, but the overloading there is not likely to cause horrendous problems (unless there are actions that are also classes of actions). A two-property solution for agents of actions would have the same characteristics. Peter F. Patel-Schneider (talk) 19:13, 14 August 2024 (UTC)[reply]
@Peter F. Patel-Schneider: I feel like you aren't hearing what I'm saying. "How then, is one to determine whether Richard Nixon (Q9588) is the kind of thing that can be the agent of a crime (Q83267)?” My response to that is, why is it a requirement that this property enable that determination? We are not creating a constraint here. Is it not enough to be able to infer Nixon's criminality from 1)the fact that he is given as the agent of a specific crime, and that 2)crimes as a class only admit criminals as agents? I don't see why it has to matter whether Richard Nixon (Q9588) is explicitly an instance of criminal (Q2159907) or not.
Now, what you call a "nasty kind of ambiguity" is the normal state of affairs for dozens if not hundreds of properties, of which I have given a few examples. It's even baked into the software: a number of constraint types take the argument relation (P2309), which takes values of instance of (Q21503252), subclass of (Q21514624), and instance or subclass of (Q30208840). If properties are only supposed to accept instances or subclasses of any particular class, then the latter value should not exist. I asked for an example of this overloading causing a specific problem, i.e. where the correct meaning is not inferable, and I still haven't seen one. To my knowledge, there are only two pairs of properties that follow this instance-version/class-version separation: has part(s) (P527)/has part(s) of the class (P2670), and object of action (P12912)/object class of action (P12913), and the separation isn't needed there either, as the correct meaning between instances and classes is eminently inferable. Swpb (talk) 23:03, 14 August 2024 (UTC)[reply]
@Swpb I'm trying to improve Wikidata and I hope that others are as well. I don't see that there is any reason to add yet another problem to Wikidata just because similar problems already exists. Peter F. Patel-Schneider (talk) 23:10, 14 August 2024 (UTC)[reply]
@Swpb I think that I am hearing what you are saying. I am not trying to scuttle this proposal because of any other reason than I think it needs to be improved. And I'm not even trying to scuttle it at all but instead improve it.
My view is that a solution for thematic relations like agent and object of actions should do two things. First it should state what kind of values are (almost) always used for the action and second it should say what the role of the value is. So the agent in a throwing action is (almost) always an object that can perform physical actions, that is an instance of some appropriate class irregardless of its participation in any action, and its role would be the thrower, but that does not make the agent be an instance of any class. Peter F. Patel-Schneider (talk) 23:24, 14 August 2024 (UTC)[reply]
@Swpb So how would you determine whether impeachment in the United States (Q1949797)agent of actionUnited States House of Representatives (Q11701) is saying that United States House of Representatives (Q11701) performs the impeachments or instances of United States House of Representatives (Q11701) perform the impeachments. Peter F. Patel-Schneider (talk) 23:51, 14 August 2024 (UTC)[reply]
And what I'm saying is, this kind of overloading is not generally a problem, either here or on other properties, and the fact that it was anticipated and allowed for in the design of the constraints indicates that that has always been the default assumption for Wikidata: some properties are constrained to only accept instance or subclass values, and others to explicitly accept either. As I said earlier, if this overloading issue is all that's holding the proposal back, I'll accept a split version, but I really don't think it's necessary, and I think those arguing for it need to present a stronger case that the overloaded version would create truly harmful ambiguity. I don't think your House of Representatives example rises to that level, since both interpretations are correct: the House, as a continuous institution, carries out impeachments, and particular sessions of the House carry out particular impeachments. So, I'm asking for a better example.
I see what you are saying about wanting to separately express a selectional restriction and a role; I'm going to take my own look at how "object class of action" is being used and see what I think we should do. I don't think relying on a qualifier is the solution, because a big use case for these properties is as qualifiers.
It seems you've added a second "oppose" vote for yourself below. You already voted "oppose" above; you should not be voting more than once. Please remove the second vote, and either move your comments under this bullet, or mark them as a comment instead of a vote. Thank you. Swpb (talk) 15:44, 15 August 2024 (UTC)[reply]
@Swpb Where are the conventions for this sort of discussion described? I found what I view as another problem in the proposal and marked it as a reason for opposition. Keeping discussion of that problem in a separate place seems reasonable to me and it also seems reasonable to me to have the discussion marked as an oppose discussion. Peter F. Patel-Schneider (talk) 16:02, 15 August 2024 (UTC)[reply]
Peter, it is fine for you to have multiple reasons to oppose the proposal and discuss them all, but it is not ok for you to vote twice. You are one person. Please adjust accordingly. Swpb (talk) 16:11, 15 August 2024 (UTC)[reply]
There is window cleaning (Q3124765)object class of action (P12913)window (Q35473) which looks like a selectional restriction, i.e., all objects of instances of window cleaning (Q3124765) are instances of window (Q35473) regardless of their participation in any action. (The force of this kind of statement might be made somewhat weaker by changing "all" to "almost all" or "normally", which would make them selectional preferences.) A large majority of the curent uses of object class of action (P12913) appear to be of this form. All three of the provided examples for object class of action (P12913) are also of this form.
But then there is magic collecting (Q110581477)object class of action (P12913)souvenir (Q212469) which looks like a role, i.e., objects of instances of magic collecting (Q110581477) are not necessarily instances of souvenir (Q212469) outside of their participation in the action. But there appears to be only a small number of these uses of object class of action (P12913). All three examples of "agent of action" on action classes appear to be of this form.
In the end, most uses of object class of action (P12913) appear to be selectional restrictions, not roles. There also appears to me to be a need for selectional restrictions on agents of actions, as agents of roles and objects of roles are very similar concepts. I also see a need for roles for agents of actions, as demonstrated in the proposal here. So there would be both crime (Q83267)agent role of actioncriminal (Q2159907) and crime (Q83267)agent class of actionperson (Q215627).
As a result I oppose this proposal until it provides a way of stating both selectional restrictions and roles for agents of actions. It might be able to do this by using a qualifier, combining both as in crime (Q83267)agent class of actionperson (Q215627) object of statement has role (P3831) criminal (Q2159907). This has the disadvantage of repeating the selectional restriction when providing a new role and having a role become a selectional restriction when editors don't figure out that a role needs a qualified statement.  – The preceding unsigned comment was added by Peter F. Patel-Schneider (talk • contribs).
I concur with your assessment that most of the objects of object class of action (P12913) are selectional restrictions, and a small minority are roles. Where I get stuck on the idea of separate properties for selectional restrictions vs. roles is that the distinction isn't always clear. Take:
I have about 10 other examples so far, where I can see different editors making different calls. The same fuzziness would apply to agents. Now, I didn't craft examples 4-6 on this proposal (I believe Wd-Ryan did), but they do suggest an asymmetry with objects of actions, where the most pertinent information about an agent will be its role more often than its selectional class. This makes sense to me, as most agents will be people or groups of people, broad classes whose members take on a lot of roles. Swpb (talk) 16:30, 15 August 2024 (UTC)[reply]

Withdrawing. Someone can redo this one that has more time than me rn to focus on WD. Lectrician1 (talk) 21:55, 15 August 2024 (UTC)[reply]

Retrieved from "https://www.wikidata.org/w/index.php?title=Wikidata:Property_proposal/agent_of_action&oldid=2228376978"