Staps & Rooryck (2024): Formalizing Spatial-Causal Polysemy of Agent Prepositions #
@cite{staps-rooryck-2024}
Semantics and Pragmatics 17, Article 4, pp. 1–47.
Core Claims #
Agent prepositions (by, French par/de) are not semantically vacuous case markers. They have non-trivial semantic content derived from their spatial meanings via principled polysemy (@cite{tyler-evans-2003}).
The general denotation is polymorphically typed: ⟨η, ⟨θ, t⟩⟩, where η (Figure) and θ (Ground) are instantiated by the syntactic/semantic context. Spatial: ⟨e, ⟨e, t⟩⟩. Agentive: ⟨e, ⟨s, t⟩⟩.
French par ('through, via') presupposes high proto-agentivity; de ('from, of') presupposes low proto-agentivity. The relevant factors are: bringing about a change (§3.1), contextually inferred change (§3.2), volitionality (§3.3), and telicity (§3.4).
The same polymorphic mechanism extends to causal adjuncts (§2.3): de marks causes-as-situations (type ⟨s, ⟨s, t⟩⟩), par marks causes-as-forces (type ⟨f, ⟨s, t⟩⟩), following @cite{copley-harley-2022}.
Formalization Strategy #
Proto-agentivity is formalized via EntailmentProfile.pAgentScore and
the specific entailments {volition, causation, movement}. The prediction
is: par requires the subject to have at least one of these "active"
entailments; de is the default for low-agentivity subjects.
We derive the par/de preference from the verb's stored EntailmentProfile
and VendlerClass, connecting to the affectedness hierarchy
(@cite{beavers-2010}) and aspectual classification (@cite{vendler-1957}).
Integration with linglib #
- Subject profiles reuse canonical @cite{dowty-1991} profiles
(
kickSubjectProfile,seeSubjectProfile,runSubjectProfile) rather than duplicating them, making the connection structural. - Passive voice connects to @cite{collins-2005}'s
voicePassive— Case-checking by Voice is the syntactic correlate of the preposition's semantic contribution. - Object affectedness derives from @cite{beavers-2010}'s
profileToDegree.
Equations
- StapsRooryck2024.instDecidableEqSemDomain x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯
Equations
- StapsRooryck2024.instReprSemDomain = { reprPrec := StapsRooryck2024.instReprSemDomain.repr }
Equations
- One or more equations did not get rendered due to their size.
Instances For
A polymorphic preposition type: ⟨η, ⟨θ, t⟩⟩. The Figure (η) is the first argument; the Ground (θ) is the second.
Instances For
Equations
- StapsRooryck2024.instDecidableEqPrepType.decEq { figure := a, ground := a_1 } { figure := b, ground := b_1 } = if h : a = b then h ▸ if h : a_1 = b_1 then h ▸ isTrue ⋯ else isFalse ⋯ else isFalse ⋯
Instances For
Equations
- StapsRooryck2024.instReprPrepType = { reprPrec := StapsRooryck2024.instReprPrepType.repr }
Equations
- One or more equations did not get rendered due to their size.
Instances For
Spatial instantiation: ⟨e, ⟨e, t⟩⟩. "The house by the lake" — two entities in close proximity.
Equations
- StapsRooryck2024.spatialType = { figure := StapsRooryck2024.SemDomain.e, ground := StapsRooryck2024.SemDomain.e }
Instances For
Agentive instantiation: ⟨e, ⟨s, t⟩⟩. "Written by John" — entity in close proximity to event. Interpretation: Initiator(x, e).
Equations
- StapsRooryck2024.agentiveType = { figure := StapsRooryck2024.SemDomain.e, ground := StapsRooryck2024.SemDomain.s }
Instances For
Causal adjunct (force-based): ⟨f, ⟨s, t⟩⟩. French par un tremblement de terre — force in close proximity to situation. Used by par in causal adjuncts.
Equations
- StapsRooryck2024.causalForceType = { figure := StapsRooryck2024.SemDomain.f, ground := StapsRooryck2024.SemDomain.s }
Instances For
Causal adjunct (situation-based): ⟨s, ⟨s, t⟩⟩. French de faim — situation in close proximity to situation. Used by de in causal adjuncts.
Equations
- StapsRooryck2024.causalSituationType = { figure := StapsRooryck2024.SemDomain.s, ground := StapsRooryck2024.SemDomain.s }
Instances For
The "active" P-Agent entailments that drive par selection. §3.5 (34): change/causation, volitionality, and kinesis/movement are the three entailments that distinguish par-requiring from de-allowing contexts. These are a strict subset of @cite{dowty-1991}'s five P-Agent entailments — sentience and independent existence do NOT contribute to par selection (experiencers allow de).
Note: the paper's (34) establishes a hierarchy among these factors: (34a) change is primary — if change is possible, par presupposes it; (34b) volitionality and telicity are secondary — relevant only when change is excluded. Our disjunction correctly predicts the binary par-required/de-available outcome for all attested verbs.
Equations
- StapsRooryck2024.hasActiveAgentEntailment p = (p.volition || p.causation || p.movement)
Instances For
par is required when the subject has active agent entailments. This is a necessary condition: de is RULED OUT for high-agentivity contexts, while par as the modern French default can appear even with low-agentivity Agents.
Instances For
de is available when active agent entailments are absent.
Instances For
§3.5 (34): par and de are in complementary distribution w.r.t. proto-agentivity. Par implies change; de presupposes its absence. This is the paper's primary empirical generalization.
Canonical agents (laver, écrire, construire, tuer, abandonner,
délaisser) have V+S+C+M+IE → par required. This is kickSubjectProfile
from @cite{dowty-1991}.
Experiencer subjects (aimer, adorer, respecter) have S+IE only →
de available. This is seeSubjectProfile from @cite{dowty-1991}.
Stative positional subjects (précéder, stative suivre) have IE only → de available.
Dynamic motion subjects (dynamic suivre) have V+S+M+IE → par
required. This is runSubjectProfile from @cite{dowty-1991}.
Accompany subjects have S+M+IE → par required (movement suffices).
§3.4: Telicity influences par/de selection. Telic events imply a bounded change, which increases proto-agentivity.
Equations
Instances For
§3.1: Object affectedness also matters. Verbs with quantized or nonquantized affectedness entail a Patient change → forces par. Connects to @cite{beavers-2010}'s affectedness hierarchy.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Full prediction combining subject proto-agentivity, object affectedness, and telicity. Any one factor suffices for par.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Par-required verbs: all have active agent entailments, affected objects, and/or telic aspect.
Par/de verbs: experiencer and stative subjects lack active agent entailments, unaffected objects, and atelic aspect.
Survey judgment: verb form, context agentivity level, and mean acceptability for de and par on [-1, 1] scale.
- verb : String
- context : String
- deMean : ℤ
- parMean : ℤ
Instances For
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Survey data from @cite{staps-rooryck-2024} Table 2. Means are ×100 (e.g., 0.87 → 87, -0.96 → -96).
Equations
- One or more equations did not get rendered due to their size.
Instances For
§3.1 Change-entailing verbs reject de (all deMean < 0). Filter restricts to the §3.1 items (high proto-agentivity context) to exclude §3.3 volitionality items with the same verb form.
High-agentivity suivi prefers par (parMean > deMean).
Low-agentivity suivi prefers de (deMean > parMean).
The by-phrase is an argument of v, not an adjunct. In the passive, Voice checks Case but does not assign θ — the θ-role comes from v. The par/de preposition is the morphological spell-out of this Case-checking relationship, with the semantic contribution being Initiator(x,e) plus the proto-agentivity presupposition.
This connects @cite{staps-rooryck-2024}'s §2.2 to @cite{collins-2005}'s passive structure: Voice_PASS checks Case while v assigns the external θ-role.
- par : FrenchAgentPrep
- de : FrenchAgentPrep
Instances For
Equations
- StapsRooryck2024.instDecidableEqFrenchAgentPrep x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Both par and de in passives have agentive type ⟨e, ⟨s, t⟩⟩. They differ only in presupposition, not at-issue type.
Equations
Instances For
In causal adjuncts, par and de diverge in type: par: ⟨f, ⟨s, t⟩⟩ (forces), de: ⟨s, ⟨s, t⟩⟩ (situations).
Equations
Instances For
The syntactic precondition for par/de: passive Voice checks Case (feature dissociation per @cite{collins-2005}), providing the structural position for the preposition. The par/de choice is then determined by proto-agentivity.
Position in the causal chain relative to the Patient. @cite{croft-2012}: par 'through' = perlative (path position), de 'from' = ablative (source position). Closer to Patient → more proto-agentive.
- source : CausalChainPosition
- path : CausalChainPosition
- goal : CausalChainPosition
Instances For
Equations
- StapsRooryck2024.instDecidableEqCausalChainPosition x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯
Equations
- One or more equations did not get rendered due to their size.
Instances For
Equations
Distance from Patient: source > path > goal.
Equations
Instances For
Equations
Instances For
Par has smaller causal distance → higher proto-agentivity.
Verbs with quantized object affectedness always require par. Quantized affectedness implies the subject causes a definite change (@cite{beavers-2010}).
Dynamic VendlerClasses (activity, achievement, accomplishment) correlate with higher proto-agentivity than states — the stative/dynamic contrast driving suivre's polysemy (§3.3).
§3.2 contextual enrichment gap: adorer's lexical profile predicts
de-availability, but high-agentivity contexts license par (survey
data shows parMean > 0 in high context). EntailmentProfile alone
cannot capture this — it requires pragmatic enrichment.
stativePositionalSubjectProfile is definitionally equal to
minimalParticipantProfile — this is structural (via abbrev),
not a coincidence. Both roles represent participants with only
independent existence.
The "active" agent entailments (V, C, M) are a strict subset of @cite{dowty-1991}'s full P-Agent set (V, S, C, M, IE). Profiles with active entailments always have pAgentScore ≥ 1, but the converse fails: experiencer profiles have pAgentScore = 2 (S+IE) yet no active entailments — which is exactly why they allow de.
Converse fails: experiencer profiles have pAgentScore ≥ 1 but no active agent entailments. This is the theoretical crux — @cite{staps-rooryck-2024}'s par/de distinction is sensitive to a narrower notion of agentivity than @cite{dowty-1991}'s full count.
Canonical agents have all three active entailments and the maximum pAgentScore of 5. This is the canonical case where @cite{staps-rooryck-2024}'s active-agent test and @cite{dowty-1991}'s full P-Agent count agree.
Proto-agentivity level #
Binary classification grounding the par/de presupposition contrast.
Derived from hasActiveAgentEntailment — not independently stipulated.
Binary proto-agentivity level for par/de presuppositions. HIGH = at least one of {V, C, M}; LOW = none.
- high : ProtoAgentivityLevel
- low : ProtoAgentivityLevel
Instances For
Equations
- StapsRooryck2024.instDecidableEqProtoAgentivityLevel x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯
Equations
Equations
- One or more equations did not get rendered due to their size.
Instances For
Bridge from EntailmentProfile to binary agentivity level.
Derived from hasActiveAgentEntailment (§2).
Equations
Instances For
Agentive denotations (35a/35b) #
The paper's key semantic contribution: par and de share at-issue content
(Initiator(x,e)) but carry complementary presuppositions about
proto-agentivity. Formalized using PrProp W from
Core.Semantics.Presupposition.
Parameters for agentive preposition denotations (35a/35b). Bundled so the denotation is parametric over the event/entity model.
- initiator : Entity → Event → W → Prop
The at-issue relation: Initiator(x, e).
- highProtoAgentivity : Entity → W → Prop
Proto-agentivity predicate over entities (in context).
Instances For
⟦par⟧_agentive (35a): λx.λe. Initiator(x,e); presup: HIGH proto-agentivity(x).
Equations
- StapsRooryck2024.parAgentiveDenot params x e = { presup := params.highProtoAgentivity x, assertion := params.initiator x e }
Instances For
⟦de⟧_agentive (35b): λx.λe. Initiator(x,e); presup: LOW proto-agentivity(x).
Equations
- StapsRooryck2024.deAgentiveDenot params x e = { presup := fun (w : W) => ¬params.highProtoAgentivity x w, assertion := params.initiator x e }
Instances For
Par and de share at-issue content — same assertion (Initiator(x,e)).
This is structural: both pass params.initiator x e as the assertion.
Par and de presuppositions are complementary: exactly one is satisfied at every world (classical).
Grounding: presuppositions derive from hasActiveAgentEntailment #
The presupposition infrastructure connects back to the per-profile
prediction machinery from §2. The proto-agentivity predicate in
AgentivePrepParams grounds out in hasActiveAgentEntailment — by
construction, not by bridge theorem.
Construct AgentivePrepParams from EntailmentProfile-based
predictions. The proto-agentivity predicate grounds out in
hasActiveAgentEntailment, connecting the formal denotation to
the existing per-profile predictions.
Equations
- One or more equations did not get rendered due to their size.
Instances For
The presupposition of ⟦par⟧_agentive agrees with predictsParRequired.
This closes the loop: the formal denotation's presupposition IS the
existing prediction, not a separate claim that happens to match.
Causal denotations (14b/15b) #
The causal uses of par and de differ in type — forces vs situations —
already captured by causalForceType/causalSituationType (§1).
The denotations are thin wrappers over the causal relations.
⟦par⟧_causal (14b): λf.λe. f FORCES e. Force-based causation — par marks the cause as a force (@cite{copley-harley-2022}).
Equations
- StapsRooryck2024.parCausalDenot Force Event W forces f e w = forces f e w
Instances For
⟦de⟧_causal (15b): λs'.λe. s' CAUSE e. Situation-based causation — de marks the cause as a situation (@cite{copley-harley-2022}).
Equations
- StapsRooryck2024.deCausalDenot Situation Event W cause s' e w = cause s' e w
Instances For
Unified dispatch #
Full denotation dispatch for agentive uses. Wraps the par/de choice
into a single function matching on FrenchAgentPrep.
Equations
- StapsRooryck2024.FrenchAgentPrep.par.agentiveDenot params x e = StapsRooryck2024.parAgentiveDenot params x e
- StapsRooryck2024.FrenchAgentPrep.de.agentiveDenot params x e = StapsRooryck2024.deAgentiveDenot params x e