Wikidata:Property proposal/has semantic role

‎has semantic role

Originally proposed at Wikidata:Property proposal/Generic

   Not done

Description	item that describes a role in an event class
Data type	Item
Domain	item, occurrence (Q1190554)
Example 1	military offensive (Q2001676)has semantic roleattacker (Q31924059)
Example 2	military offensive (Q2001676)has semantic roledefender (Q111729140)
Example 3	throwing (Q12898216)has semantic roleactor (Q23894381)
Example 4	throwing (Q12898216)has semantic roletarget (Q1047579)
Example 5	throwing (Q12898216)has semantic roleprojectile (Q49393)
Planned use	add to (possibly newly created) items describing occurrences/actions
Expected completeness	always incomplete (Q21873886)

This proposal replaces the following property proposals:

event role, role in event, selectional preference, event argument and argument type

Motivation

All eventualities, including events, states, actions and processes, have core semantic roles, as illustrated by widely used resources such as FrameNet, VerbNet and PropBank. “Eating" has an "eater" and something "eaten"; "throwing" has the "thrower", the "target" and the "projectile". These roles are not optional. Every act of "eating" has an "eater" and something "eaten" independently of how it is expressed and in what language. While Wikidata has over 300 existing properties for roles in event instances (e.g., participant (P710), victim (P8032), there are very few that are used with event/process classes. The two most common are practiced by (P3095) and uses (P2283). The vast majority of event/process classes have no statements describing semantic roles. Our proposed “has semantic role” property is designed to fill this gap. Existing properties like practiced by (P3095), which is used with items such as eater (Q20984678), should become a subproperty of (P1647) of “has semantic role”. The Wikidata item of this property (P1629) will be semantic role (Q117747915) which currently does not have a corresponding property. We do not want to duplicate information that is already present, but rather integrate within a coherent, consistent overarching framework.

The object of “has semantic role” property will be an item that describes the role. Whenever possible, we will find an existing item such as eater (Q20984678), otherwise, we will create a new item. We also want to provide a broad characterization of the type of role it is, such as Actor or Undergoer, as explained in the Semantic Roles subsection below.

We will use PropBank (Q7250039), the largest repository of structured event, process and action descriptions (over 11,000 role sets) to identify the existing or missing event/process items and add their semantic role statements.

For instance, Item work (Q268378) can map to the PropBank work.01 roleset. The item has a statement work (Q268378)has part(s) (P527)task (Q759676) which is similar to PropBank’s “job, project”, Theme (ARG1-PPT). PropBank also lists several other roles that are missing from the Wikidata item. Of these, the ‘worker’ or Actor role (ARG0-PAG) should certainly be added, and probably the employer (Affectee, ARGX-GOL). See the subsection on Semantic Roles for definitions of Actor and Undergoer. We could either use the generic “has semantic role” proposed property, or any of the properties we have identified below as potential subproperties of “has semantic role” (i.e., adding statements such as work (Q268378)practiced by (P3095)worker (Q327055) and work (Q268378)has characteristic (P1552)employer (Q3053337).

Integrating “has semantic role” with existing properties

Wikidata has several properties that already highlight critical semantic relations between eventualities and their participants. We can relate the existing properties such as practiced by (P3095) to "has semantic role" using statements such as practiced by (P3095)subproperty of (P1647)has semantic role. We envision “has semantic role” as the most general, and therefore the top of a simple hierarchy that would include, at a minimum, the following properties: practiced by (P3095), uses (P2283), has characteristic (P1552), has part(s) (P527), has cause (P828), has effect (P1542). (These properties can have other subproperty of (P1647) statements as well.)

We have done some manual inspection of the current usage of the above properties. practiced by (P3095) can typically be relied upon to describe the Actor of an eventuality, i.e., {{Statement|Q213449|P3095|Q20984678)” and marketing (Q39809)practiced by (P3095)marketer (Q1900657).

Item uses (P2283) also describes instruments associated with eventualities coloring (Q2022532)uses (P2283)colourant (Q911922), or grinding (Q26882416)uses (P2283)mill (Q44494) but can also be much more variable, for exa mple, sometimes describing an Undergoer relationship eating (Q213449)uses (P2283)food (Q2095) or even an Actor relationship transport (Q7590)uses (P2283)agent (Q24229398).

The remaining properties describe semantic roles variously; budget (Q41263)has characteristic (P1552)budget constraint (Q605095) and marketing (Q39809)has part(s) (P527)product (Q2424752) which both describe Undergoer relationships.

These properties can all describe semantic roles, but which roles they describe can sometimes be ambiguous. We propose using an existing WD qualifier object has role (P3831) to make such roles unambiguous, using PropBank as our guide, e.g., budget (Q41263)has characteristic (P1552)budget constraint (Q605095)object has role (P3831)undergoer (Q111335542).

Semantic Roles

When no value exists for a given qualifier (e.g., worker for the actor of a working event), a set of semantic roles (e.g., actor (Q23894381), undergoer (Q111335542) will be used. Below is a table listing these semantic roles, adopted from the Uniform Meaning Representation project and used by PropBank, which have been carefully reviewed to ensure that they accommodate cross-linguistic typological variation (Bonial et al. 2011 A Hierarchical Unification of LIRICS and VerbNet Semantic Roles (Q118174236), Van Gysel et al, 2021 Designing a Uniform Meaning Representation for Natural Language Processing (Q115519832)). For the most part we will be relying on existing Wikidata Thematic Relation definitions to realize our PropBank semantic roles, as illustrated in this table. It shows how existing items will be used and ensures forwards and backwards compatibility with no disruption to Wikidata structures that may already be in use. This systematic approach also ensures that future items added to WD will be able to utilize this system to benefit from the enriched event representation.

Semantic Roles

Semantic Role	Wikidata item	Semantic Role	Wikidata item	Semantic Role	Wikidata item
Actor	actor (Q23894381)	instrument	instrument (Q6535309)	Cause	cause (Q2574811)
Causer	agent (Q392648)	Start	origin (Q3885844)	Temporal	duration (Q2199864) time (Q12322185) Frequency (Q125995799)
Force	force (Q126009669)	Goal	goal (Q109405570)	Extent	extent (Q125953445)
Undergoer	undergoer (Q111335542) patient (Q170212)	Companion	companion (Q106645134)	Manner	means (Q12774177)
Theme	theme (Q118826633)	Material/Source	material (Q214609) source (Q31464082)	Reason	cause (Q2574811)
Recipient	recipient (Q20820253) addressee (Q19720921)	Place	location (Q109377685)	Purpose	cause (Q2574811)
Experiencer	experiencer (Q1242505)	Affectee	affectee (Q125995757)	Attribute	attribute (Q109674924)
Stimulus	stimulus (Q109566760)	Direction	direction (Q2151613)	Result	result (Q2995644)

Anatole Gershman (talk) 17:01, 24 May 2024 (UTC)

Discussion

•  Comment @ChristianKl, Peter F. Patel-Schneider, Arademaker, Swpb, ArthurPSmith: from the previous proposals. Mahir256 (talk) 17:29, 24 May 2024 (UTC)

  To get the easiest out of the way first, what does the German town of Mittenwalde (Q574811) have to do with "Purpose"?

  The description speaks of "event class" but neither throwing (Q12898216) nor military offensive (Q2001676) are events in our ontology. Given that I made that point a few times already, why are you still talking about event classes? Event class is a term foreign for Wikidata and also not everyday language where you can expect that everyone will understand it the same way.

  It's unclear to me what the word semantic does here. Items are not words or their labels. Labels are semantic objects and point to concepts (items). If you actually want to speak about semantic entities, we have lexemes. throw (L28480) is a semantic entity. I would see less of an issue if this proposal would switch to focus on lexemes instead of focusing on items.

  throwing (Q12898216)has semantic roleactor (Q23894381) seems to be a good example of why the proposal is problematic. It somehow tries to store information about who does the throwing but it doesn't let us know that pitcher (Q1048902) is someone who throws. So in total I  Oppose this proposal as well. ChristianKl ❪✉❫ 21:16, 24 May 2024 (UTC)

  I'm not so worried about "event" as the current proposal includes actions. Just replace that wording with something better, perhaps occurence. Peter F. Patel-Schneider (talk) 15:41, 26 May 2024 (UTC)

•  Comment Also see this list of actions that don't have specific parent classes: User:Wd-Ryan/Basic_actions. A lot of work to be done to model actions on Wikidata. -wd-Ryan (Talk/Edits) 03:36, 26 May 2024 (UTC)
•  Comment The proposal needs complete examples to see how it would work, including not just the information associated with event classes but also information associated with event instances. Without such examples it is hard to determine just what the proposal involves.

For example, the proposal appears to indicate that the information added to the military offensive (Q2001676) action class contains two values for "has semantic role", namely attacker (Q31924059) and defender (Q111729140). But how does this impact information on instances of military offensive (Q2001676), such as Petsamo–Kirkenes Offensive (Q705222), which has two values for participant (P710)? The proposal seems to indicate that this is done (for existing properties) by making them subproperties of "has semantic role". The proposal then appears to go on and say that there should be two values for participant (P710) on military offensive (Q2001676).

But this is bad modelling. Properties like participant (P710) are for individual actions like Petsamo–Kirkenes Offensive (Q705222) not classes like military offensive (Q2001676). The relationship between participant (P710) and "has semantic role" is not subproperty. So the method of integrating existing properties with "has semantic role" is flawed and something different needs to be done, perhaps making the values of "has semantic role" be properties.

The proposal is also silent on how it should work in a clean state. Consider again Petsamo–Kirkenes Offensive (Q705222). What should be added to it and to military offensive (Q2001676) and possibly to other items to convey the information that the attacker in this action is Soviet Union (Q15180) and the defender is Nazi Germany (Q7318) if there were no relevant existing properties in Wikidata? How this is to be done needs to be shown for the proposal to be acceptable.

So  Oppose unless these points are addressed. Peter F. Patel-Schneider (talk) 15:39, 26 May 2024 (UTC)

military offensive (Q2001676) properties for this type (P1963) participant (P710) is the existing syntax to say that Petsamo–Kirkenes Offensive (Q705222) should have participant (P710) statements. ChristianKl ❪✉❫ 19:58, 26 May 2024 (UTC)

Indeed. A better version of "has semantic role" would make it a subproperty of properties for this type (P1963) and change the values from classes to properties. Peter F. Patel-Schneider (talk) 00:43, 27 May 2024 (UTC)

How would it then differ from properties for this type (P1963)? ChristianKl ❪✉❫ 09:57, 27 May 2024 (UTC)

It would specialize properties for this type (P1963). properties for this type (P1963) is for any property that should normally have a value for instances of a class. "has semantic role" would be for properties that are semantic roles and that have to have a value for instances of an event/action/... (but the value or values might not be present in Wikidata). But maybe this is too fine a distinction and properties for this type (P1963) is what should be used to signal semantic roles. Peter F. Patel-Schneider (talk) 13:52, 27 May 2024 (UTC)

This is an important point that we discussed in our previous proposal but should have addressed here. The object of the proposed property is an item that describes the role. It is not a role filler itself. In the example of military offensive (Q2001676), the item attacker (Q31924059) describes the attacker role. The actual attacker in an instance of a military offensive does not have to descend from attacker (Q31924059). We could add statements to attacker (Q31924059) that specify selectional preferences for the role fillers (e.g., that they should descend from military (Q8473)). Petsamo–Kirkenes Offensive (Q705222) is an instance of military offensive (Q2001676) and it has two participant (P710) statements with objects: Soviet Union (Q15180) and Nazi Germany (Q7318) without specifying which participant was the attacker and which was the defender. We propose to add the "object has role (P3831)" qualifier to these statements: "attacker (Q31924059)participant (P710)Soviet Union (Q15180)object has role (P3831)attacker (Q31924059)" and "attacker (Q31924059)participant (P710)Nazi Germany (Q7318)object has role (P3831)defender (Q111729140)". As you correctly stated, participant (P710) is for instances while "has semantic role" is for classes. We do not propose to subordinate participant (P710) to "has semantic role". I hope this clarification helps. Anatole Gershman (talk) 22:28, 30 May 2024 (UTC)

Your comment appears to contradict `We can relate the existing properties such as practiced by (P3095) to "has semantic role" using statements such as practiced by (P3095) subproperty of (P1647) has semantic role' so I am now confused. Fully worked-out examples would help (provided that they match the proposal wording) and object has role (P3831) is not in the proposal so at the very least this important facet needs to be included in the proposal itself. These examples should say what is and what is not allowed. For example, can any statement on Petsamo–Kirkenes Offensive (Q705222) be annotated with object has role (P3831) no matter the type of the value? Peter F. Patel-Schneider (talk) 23:45, 30 May 2024 (UTC)

In our analysis, "practiced by (P3095)" is one of the very few properties that are used with event/action classes. So, indeed, it would become a subproperty of (P1647) "has semantic role" in our proposal. Another example is "uses (P2283)". There are 300+ properties (including participant (P710)) that are used exclusively with event/action instances and whose objects are the actual role fillers. We do not propose to subordinate them to "has semantic role", but we suggest using the "object has role (P3831)" qualifier to indicate the role their objects are playing. You are right that we should include the whole example and the use of the "object has role (P3831)" qualifier in the main body of the proposal. We will do that. On the "what's allowed and what is not" question, we can only offer guidelines. Thank you. Anatole Gershman (talk) 00:45, 31 May 2024 (UTC)

OK, there are only a few properties that you will make suproperties of "has semantic role". Do you have an exhaustive list? One issue is that you say that one of these properties is has characteristic (P1552) but this property seems to be much more general than "has semantic role", not least in that it appears to be relevant for any class, not just events/actions/.... How can you resolve this inconsistency? uses (P2283) appears to have the same problem.

You are proposing quite a large addition to how events/actions/... are to be modelled in Wikidata so I think that there needs to be more than guidelines, but perhaps not inviolable rules. For example, Petsamo–Kirkenes Offensive (Q705222) has values for participant (P710), which are suitable for object has role (P3831) qualifiers. It also has values for country (P17), start time (P580), end time (P582), and part of (P361). Which, if any, of these are suitable for object has role (P3831) qualifiers? Further, it seems that start time (P580) and end time (P582) are actually semantic roles for Petsamo–Kirkenes Offensive (Q705222). How are these two properties, and other similar properties, going to relate to your proposal? A fully worked-out example would show how all this works. Peter F. Patel-Schneider (talk) 02:52, 31 May 2024 (UTC)

Thank you again for insightful comments. We do not have an exhaustive list of current properties that are used with events/actions to indicate semantic roles. This is an ongoing process as we map the 11,000+ PropBank concepts to Wikidata. For example, as we mapped 100 most frequent English verbs to the underlying concepts in PropBank and Wikidata, we found 6 properties used to indicate the "core" semantic roles as defined in PropBank. These are: "practiced by (P3095)", "uses (P2283)", "has effect (P1542)", "immediate cause of (P1536)", "participant (P710)", and "location (P276)". Of these, only the first was used extensively and the rest occasionally. One might argue about the appropriateness of some of these uses but we don't have any control of how people use the existing properties leading to inevitable inconsistencies. Because of this, we need to be judicious in making the existing properties sub-properties of "has semantic role". For example, "practiced by (P3095)" would be a good candidate but, as you noted, "has characteristic (P1552)" and "location (P276)" would be not. We will revise the paragraph dealing with the "has semantic role" sub-properties.

If we don't make "uses (P2283)" a sub-property of "has semantic role", how would one know that "food (Q2095)" in "eating (Q213449)uses (P2283)food (Q2095)" is a semantic role? The table in our proposal suggests that "food (Q2095)" should be a descendant of "semantic role (Q117747915)" via "theme (Q118826633)".

You are right that event/action classes and their instances can have many semantic roles. Currently, we are focusing only on the "core" roles as defined in PropBank. Any semantic role of an event/action instance can have a "object has role (P3831)" qualifier that would indicate the type of role the object is playing.

You are also right that we are proposing extensive additions to the event/action classes in Wikidata. This is because the current coverage of these classes in Wikidata is very sparse. We are proposing to use PropBank as a source to systematically add missing data. These additions can be done gradually without changing any of the existing statements. Anatole Gershman (talk) 14:50, 3 June 2024 (UTC)

You say that practiced by (P3095) is a good candidate to be a subproperty of "has semantic role". But practiced by (P3095) has given examples like volcanology (Q102904) practiced by (P3095) volcanologist (Q7940086). volcanology (Q102904) is not part of the domain of "has semantic role" so it is not a candidate to be a subproperty of "has semantic role". This indicates that your analysis of Wikidata is incorrect and needs to be redone. Similarly, if you make uses (P2283) a subproperty of "has semantic role" then all statements using uses (P2283) will also be statements about "has semantic role", so 1983 Belgian Grand Prix (Q20920) "has semantic role" Spa-Francorchamps modern circuit with original Bus Stop Chicane (Q66436621). This does not seem to be correct.

I think that there needs to be much more description of the larger proposal of integrating PropBank and Wikidata and how this property fits into that proposal. The larger proposal should include information on what existing Wikidata properties will be involved and how they will fit into the integration. The information should include what happens with properties that are generalizations of semantic roles, like participant (P710), and what happens with Wikidata properties that appear to already be specific semantic parts of actions, like location (P276) and start time (P580).

You also say that 'Any semantic role of an event/action instance can have a "object has role (P3831)" qualifier that would indicate the type of role the object is playing.' But what makes something a semantic role of an event/action? Guidance is needed on this point. Peter F. Patel-Schneider (talk) 16:10, 4 June 2024 (UTC)

•  CommentThis proposal appears to define what types/classes of instances should be associated to a particular event class. E.g., military offensive has an agent (Q24229398) acting as the attacker and another as the defender. However, this can also be specified using SHACL or ShEx (especially as an EntitySchema).

Being defined at the class level, the current proposal says nothing about how to specify the actual agent that is the attacker or defender in a particular offensive. I.e., how does this aid in understanding a particular event instance and its related entities? This seems the real purpose of semantic roles. "Joe threw the ball to John." => Throw is the event; Joe is the actor/agent; John is the recipient; the ball is the 'theme' (what is thrown).

In addition, it is incorrect to say ... "The Wikidata item of this property (P1629) will be semantic role (Q117747915) which currently does not have a corresponding property." It does have a property ... As a subclass of role, semantic role is valid as the value of the subject has role (P2868) property.

Lastly, there are standard semantic roles (agent, experiencer, causer, ...) as noted in the table above and in various online sources (e.g., https://glossary.sil.org/term/semantic-role or https://schemantra.com/blog/2023/07/28/semantic-roles/). It may be better to explicitly capture these (has agent, has experiencer, ...) as properties for an event. This makes it easy to define the individuals in specific roles in a specific event instance. And, it reduces the need to explicitly create unique roles to distinguish attacker vs thrower in military offensive vs throwing events. The attacker or the thrower are the active actor/agent in the events.

Without clarifying the proposal to address these issues, I  Oppose it. Andrea Westerinen

@AWesterinen: subject has role (P2868) does not have a type constraint for "role". The purpose for which that property was created was to be used as a qualifier and in that main purpose A type constraint for role is not useful. In the current usage, it likely would have never made it through the property proposal process and the current usage of the property is a good illustration of the mess that this produces. Hopefully, we will get the property back to it's intented purpose sooner or later. ChristianKl ❪✉❫ 22:02, 2 June 2024 (UTC)

subject has role (P2868) has not been created to be used as a qualifier. Please have a look at the property proposal: Wikidata:Property_proposal/has_role. The discussion occured later, after the creation of object has role (P3831). You may argue that it would be better to use it only as a qualifier, but to say that making it a qualifier-only-property restores its original purpose is wrong. - Valentina.Anitnelav (talk) 16:00, 3 June 2024 (UTC)

• I'm sorry, but this will be closed as  Not done. There're too many issues and it appears in total as unusable. In addition, there is no recognizable consensus for the creation. --Wüstenspringmaus _talk 11:30, 6 June 2024 (UTC)

  It would be very useful to have a facility like this so that Wikidata information about events and actions can be connected to linguistic descriptions of events and actions. What is a good way of facilitating a discussion on an acceptable way of adding such a facility? Peter F. Patel-Schneider (talk) 12:00, 6 June 2024 (UTC)

To create such a property, we need to solve the problems mentioned above, the property must be useful in all cases of its "application area". If these points are not sufficiently represented (if they can even all be satisfied) and there is no consensus, I'm not allowed to create the property. Regards, --Wüstenspringmaus _talk 11:45, 7 June 2024 (UTC)

Yes, an acceptable proposal has to be better, but is there a process for creating and discussing a proposal of this sort that doesn't have to fit into the boundaries of a property proposal and will also be visible to interested parties? Peter F. Patel-Schneider (talk) 11:50, 7 June 2024 (UTC)