@cite{martin-schaefer-kastner-2025} — The Lexical Pragmatics of Reflexive Marking

Martin, Fabienne, Florian Schäfer & Itamar Kastner. 2025. The lexical pragmatics of reflexive marking. Language 101(3): 524–571.

Core thesis

French anticausatives marked with se or left unmarked do not differ in meaning. Rather, cooperative speakers manage the voice ambiguity introduced by se (which can mark both anticausative and reflexive voice) in line with the Manner supermaxim "Be perspicuous." The choice between ±se is driven by three interacting factors:

1. Verb class (limited-control vs. in-control)
2. Animacy of the sole DP argument (human vs. nonhuman)
3. Agent bias — the tendency to interpret human DPs as agents

Three generalizations (Table 1)

• Unmarked limited-control preference (human DP): limited-control ±se verbs (rougir 'blush') prefer the −se form. Rationale: −se is unambiguously anticausative; +se introduces a reflexive parse that clashes with shared assumptions (the change is not under the human's control). Avoiding ambiguity is optimal.

• Marked in-control preference (human DP): in-control ±se verbs (plier 'bend') prefer the +se form. Rationale: choosing −se would trigger a 'no-agent' inference (the speaker avoided the ambiguous form, signaling no agentive construal), which clashes with shared assumptions (the change IS typically under the human's control). Maintaining ambiguity is optimal.

• Marked responsibility preference (nonhuman DP): the +se form is preferred when the speaker aims to present the nonhuman as responsible. Only the se-marked form allows a reflexive parse, which is the only grammatical way to assign agency to a nonhuman sole argument.

Bridges

• VoiceFlavor.nonThematic (@cite{schaefer-2008}): the anticausative voice flavor — contributes no semantics. Both ±se anticausatives have this.
• VoiceFlavor.reflexive: the reflexive voice flavor — assigns agent + theme to the sole DP. Only available with se.
• MannerSubmaxim.avoidAmbiguity: the Manner sub-maxim driving the unmarked limited-control preference.
• cosSubjectProfile / motionCosSubjectProfile: entailment profiles for anticausative subjects. ControlLevel cross-cuts these: limited-control and in-control property-change verbs share the same profile.

Relationship to @cite{koontz-garboden-2009}

MSK2025 presupposes K-G 2009's reflexive-anticausative syncretism: se marks both anticausative and reflexive voice because anticausativization IS reflexivization. The pragmatic effects arise precisely because this syncretism creates voice ambiguity that speakers must manage.

inductive MartinSchaeferKastner2025.SeMarking : Type

Morphological class of an anticausative with respect to se-marking. @cite{martin-schaefer-kastner-2025} §1:

• −se: AC form cannot have se (changer de position)
• +se: AC form must have se (s'affaiblir)
• ±se: AC form optionally has se (casser, plier, rougir)

• minusSe : SeMarking
• plusSe : SeMarking
• plusMinusSe : SeMarking

@[implicit_reducible]

instance MartinSchaeferKastner2025.instDecidableEqSeMarking : DecidableEq SeMarking

Equations

• MartinSchaeferKastner2025.instDecidableEqSeMarking x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

@[implicit_reducible]

instance MartinSchaeferKastner2025.instReprSeMarking : Repr SeMarking

Equations

• MartinSchaeferKastner2025.instReprSeMarking = { reprPrec := MartinSchaeferKastner2025.instReprSeMarking.repr }

def MartinSchaeferKastner2025.instReprSeMarking.repr : SeMarking → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

inductive MartinSchaeferKastner2025.ControlLevel : Type

Lexical-semantic class of a change-of-state verb based on whether the change is typically under the human undergoer's control. @cite{martin-schaefer-kastner-2025} §1.1.

This classification reflects shared world knowledge about human agency, NOT lexical-semantic structure. Property-change verbs from both classes share the same entailment profile (cosSubjectProfile); the distinction is pragmatic (see control_level_not_from_entailments).

• limitedControl : ControlLevel

  Change typically NOT under human control: rougir 'blush', pâlir 'get pale', rajeunir 'rejuvenate'.

• inControl : ControlLevel

  Change typically under human control: plier 'bend', (s')approcher 'get close', (se) courber 'bend/curve'.

@[implicit_reducible]

instance MartinSchaeferKastner2025.instDecidableEqControlLevel : DecidableEq ControlLevel

Equations

• MartinSchaeferKastner2025.instDecidableEqControlLevel x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

def MartinSchaeferKastner2025.instReprControlLevel.repr : ControlLevel → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

@[implicit_reducible]

instance MartinSchaeferKastner2025.instReprControlLevel : Repr ControlLevel

Equations

• MartinSchaeferKastner2025.instReprControlLevel = { reprPrec := MartinSchaeferKastner2025.instReprControlLevel.repr }

inductive MartinSchaeferKastner2025.Animacy : Type

Animacy of the sole DP argument.

• human : Animacy
• nonhuman : Animacy

@[implicit_reducible]

instance MartinSchaeferKastner2025.instDecidableEqAnimacy : DecidableEq Animacy

Equations

• MartinSchaeferKastner2025.instDecidableEqAnimacy x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

def MartinSchaeferKastner2025.instReprAnimacy.repr : Animacy → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

@[implicit_reducible]

instance MartinSchaeferKastner2025.instReprAnimacy : Repr Animacy

Equations

• MartinSchaeferKastner2025.instReprAnimacy = { reprPrec := MartinSchaeferKastner2025.instReprAnimacy.repr }

inductive MartinSchaeferKastner2025.ResponsibilityGoal : Type

Speaker's communicative goal regarding agency attribution. G3 (responsibility preference) is conditional on this goal: with nonhuman DPs, +se is preferred only when the speaker aims to present the entity as responsible.

• neutral : ResponsibilityGoal
• conveyResponsibility : ResponsibilityGoal

@[implicit_reducible]

instance MartinSchaeferKastner2025.instDecidableEqResponsibilityGoal : DecidableEq ResponsibilityGoal

Equations

• MartinSchaeferKastner2025.instDecidableEqResponsibilityGoal x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

@[implicit_reducible]

instance MartinSchaeferKastner2025.instReprResponsibilityGoal : Repr ResponsibilityGoal

Equations

• MartinSchaeferKastner2025.instReprResponsibilityGoal = { reprPrec := MartinSchaeferKastner2025.instReprResponsibilityGoal.repr }

def MartinSchaeferKastner2025.instReprResponsibilityGoal.repr : ResponsibilityGoal → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

theorem MartinSchaeferKastner2025.control_level_not_from_entailments : Fragments.French.Predicates.rougir.subjectEntailments = Fragments.French.Predicates.refroidir.subjectEntailments

ControlLevel is NOT derivable from Dowty's entailment profiles. Limited-control property-change verbs (rougir) and in-control property-change verbs (refroidir) share the same entailment profile: both lack volition, sentience, causation, and movement.

The classification reflects shared world knowledge:

• rougir 'blush': blushing is not typically under human control
• refroidir 'cool down': cooling can be under human control

Both are non-volitional changes of state with no movement.

theorem MartinSchaeferKastner2025.movement_sufficient_not_necessary : Fragments.French.Predicates.approcher.subjectEntailments = some Fragments.French.Predicates.motionCosSubjectProfile ∧ Fragments.French.Predicates.refroidir.subjectEntailments = some Fragments.French.Predicates.cosSubjectProfile ∧ Fragments.French.Predicates.motionCosSubjectProfile.movement = true ∧ Fragments.French.Predicates.cosSubjectProfile.movement = false

Movement in the entailment profile is a SUFFICIENT but not NECESSARY condition for in-control status. Motion verbs like approcher have movement AND are in-control. But property-change verbs like refroidir are in-control WITHOUT movement.

def MartinSchaeferKastner2025.bareVoiceOptions : List Minimalist.VoiceFlavor

Voice flavors available for a form WITHOUT se. Only the anticausative (non-thematic) parse is available. The DP is a theme, not an agent.

Equations

• MartinSchaeferKastner2025.bareVoiceOptions = [Minimalist.VoiceFlavor.nonThematic]

def MartinSchaeferKastner2025.seVoiceOptions : List Minimalist.VoiceFlavor

Voice flavors available for a form WITH se. Both anticausative and reflexive parses are available — this is the reflexive-anticausative syncretism (@cite{schaefer-2008}, @cite{koontz-garboden-2009}).

K-G 2009's reflexivization analysis predicts exactly this syncretism: se marks reflexivization, which covers both the reflexive reading (the entity acts on itself) and the anticausative reading (EFFECTOR = THEME, agentivity bleached by underspecification).

Equations

• MartinSchaeferKastner2025.seVoiceOptions = [Minimalist.VoiceFlavor.nonThematic, Minimalist.VoiceFlavor.reflexive]

theorem MartinSchaeferKastner2025.se_is_ambiguous : seVoiceOptions.length > bareVoiceOptions.length

The se-marked form is ambiguous: it has strictly more voice parses.

theorem MartinSchaeferKastner2025.bare_is_unambiguous : bareVoiceOptions = [Minimalist.VoiceFlavor.nonThematic]

The bare form is unambiguous for anticausative voice.

theorem MartinSchaeferKastner2025.shared_anticausative_parse : Minimalist.VoiceFlavor.nonThematic ∈ bareVoiceOptions ∧ Minimalist.VoiceFlavor.nonThematic ∈ seVoiceOptions

Both forms share the anticausative parse.

theorem MartinSchaeferKastner2025.reflexive_only_with_se : Minimalist.VoiceFlavor.reflexive ∈ seVoiceOptions ∧ Minimalist.VoiceFlavor.reflexive ∉ bareVoiceOptions

The reflexive parse is available only with se.

def MartinSchaeferKastner2025.reflexiveParseSalient (anim : Animacy) : Bool

Agent bias: human DPs in subject position are preferentially interpreted as agents (@cite{bickel-etal-2015}, @cite{sauppe-etal-2023}). With nonhuman DPs, the reflexive parse is not a priori salient.

Equations

• MartinSchaeferKastner2025.reflexiveParseSalient MartinSchaeferKastner2025.Animacy.human = true
• MartinSchaeferKastner2025.reflexiveParseSalient MartinSchaeferKastner2025.Animacy.nonhuman = false

def MartinSchaeferKastner2025.reflexiveParseClashes (ctrl : ControlLevel) : Bool

Whether the non-target sense (reflexive) clashes with shared assumptions about the event.

• Limited-control verbs: the reflexive parse (DP = agent of own change) clashes with the assumption that the change is not under the human's control. Reflexive parse = misleading.
• In-control verbs: the reflexive parse aligns with the assumption that the change IS under the human's control. Reflexive parse = not misleading.

Equations

• MartinSchaeferKastner2025.reflexiveParseClashes MartinSchaeferKastner2025.ControlLevel.limitedControl = true
• MartinSchaeferKastner2025.reflexiveParseClashes MartinSchaeferKastner2025.ControlLevel.inControl = false

def MartinSchaeferKastner2025.predictSePreference (ctrl : ControlLevel) (anim : Animacy) : Option Bool

The predicted preference for the form of ±se AC-verbs. Returns some true if +se is preferred, some false if −se is preferred, none if no preference arises.

Derived compositionally from the paper's §2.2 reasoning:

1. Is the reflexive parse salient? (reflexiveParseSalient) Agent bias activates with human DPs, making the reflexive parse of se a priori salient. With nonhuman DPs, no ambiguity to manage.

2. If salient, does it clash with shared assumptions? (reflexiveParseClashes)

   • YES (limited-control): reflexive parse misleading → AVOID ambiguity → prefer bare (−se). @cite{dowty-1980}: if structure A is ambiguous between X and Y while B has only X, reserve A for Y.
   • NO (in-control): bare form's anti-implicature ("no agent") clashes instead → MAINTAIN ambiguity → prefer +se.

§2.3 reframes "Avoid ambiguity" as "MANAGE ambiguity" — a single Manner principle with two optimal strategies depending on whether the nontarget sense aligns with or clashes against shared assumptions.

For nonhuman DPs, see predictSePreferenceExt which incorporates the speaker's responsibility goal (G3).

Equations

• One or more equations did not get rendered due to their size.

def MartinSchaeferKastner2025.predictSePreferenceExt (ctrl : ControlLevel) (anim : Animacy) (goal : ResponsibilityGoal) : Option Bool

Extended preference prediction incorporating the speaker's communicative goal (G3: responsibility preference for nonhuman DPs).

With nonhuman DPs in neutral contexts, neither form is preferred (agent bias inactive). But when the speaker aims to convey that the nonhuman entity is responsible for the change, +se is preferred: only the se-marked form allows a reflexive parse, which is the only grammatical way to assign agency to a nonhuman sole argument.

Equations

• MartinSchaeferKastner2025.predictSePreferenceExt ctrl MartinSchaeferKastner2025.Animacy.human goal = MartinSchaeferKastner2025.predictSePreference ctrl MartinSchaeferKastner2025.Animacy.human
• MartinSchaeferKastner2025.predictSePreferenceExt ctrl MartinSchaeferKastner2025.Animacy.nonhuman MartinSchaeferKastner2025.ResponsibilityGoal.neutral = none
• MartinSchaeferKastner2025.predictSePreferenceExt ctrl MartinSchaeferKastner2025.Animacy.nonhuman MartinSchaeferKastner2025.ResponsibilityGoal.conveyResponsibility = some true

def MartinSchaeferKastner2025.preferenceRationale (ctrl : ControlLevel) (anim : Animacy) : Option Pragmatics.GriceanMaxims.MannerSubmaxim

The pragmatic principle behind each preference.

Equations

• One or more equations did not get rendered due to their size.
• MartinSchaeferKastner2025.preferenceRationale ctrl MartinSchaeferKastner2025.Animacy.nonhuman = none

theorem MartinSchaeferKastner2025.unmarked_limited_control_preference : predictSePreference ControlLevel.limitedControl Animacy.human = some false

Generalization 1 — Unmarked limited-control preference (human DP). With a limited-control ±se verb and a human DP, the −se form is preferred (experiment 1a: neutral/inchoative contexts).

theorem MartinSchaeferKastner2025.marked_in_control_preference : predictSePreference ControlLevel.inControl Animacy.human = some true

Generalization 2 — Marked in-control preference (human DP). With an in-control ±se verb and a human DP, the +se form is preferred (experiment 1b: all three contexts).

theorem MartinSchaeferKastner2025.marked_responsibility_preference : predictSePreferenceExt ControlLevel.limitedControl Animacy.nonhuman ResponsibilityGoal.conveyResponsibility = some true ∧ predictSePreferenceExt ControlLevel.inControl Animacy.nonhuman ResponsibilityGoal.conveyResponsibility = some true

Generalization 3 — Marked responsibility preference (nonhuman DP). With a nonhuman DP, +se is preferred when the speaker aims to present the entity as responsible. Only se allows the reflexive parse, which is the only grammatical way to assign agency to a nonhuman sole argument. Control level is irrelevant — the mechanism (reflexive parse as sole agency channel) applies uniformly.

theorem MartinSchaeferKastner2025.no_preference_nonhuman_neutral : predictSePreferenceExt ControlLevel.limitedControl Animacy.nonhuman ResponsibilityGoal.neutral = none ∧ predictSePreferenceExt ControlLevel.inControl Animacy.nonhuman ResponsibilityGoal.neutral = none

In neutral contexts with nonhuman DPs, no preference is predicted (agent bias inactive, no pragmatic pressure).

def MartinSchaeferKastner2025.availableForms (marking : SeMarking) : ℕ

Pragmatic form preferences arise ONLY with ±se verbs, where the speaker has a genuine choice between bare and se-marked forms.

• −se verbs: only the bare form exists → no choice, no pragmatics
• +se verbs: only the se form exists → no choice, no pragmatics
• ±se verbs: both forms available → speaker must manage ambiguity

The number of available forms determines whether pragmatic pressure applies.

Equations

• MartinSchaeferKastner2025.availableForms MartinSchaeferKastner2025.SeMarking.minusSe = 1
• MartinSchaeferKastner2025.availableForms MartinSchaeferKastner2025.SeMarking.plusSe = 1
• MartinSchaeferKastner2025.availableForms MartinSchaeferKastner2025.SeMarking.plusMinusSe = 2

theorem MartinSchaeferKastner2025.choice_only_with_plusMinusSe (m : SeMarking) : availableForms m > 1 ↔ m = SeMarking.plusMinusSe

Only ±se verbs have a genuine choice between forms.

theorem MartinSchaeferKastner2025.cos_profile_unaccusative : Semantics.ArgumentStructure.EntailmentProfile.PredictsUnaccusative Fragments.French.Predicates.cosSubjectProfile

All anticausative verb profiles predict unaccusativity (no volition, no causation, has P-Patient features). This is expected: anticausative subjects are derived internal arguments.

theorem MartinSchaeferKastner2025.motion_cos_profile_unaccusative : Semantics.ArgumentStructure.EntailmentProfile.PredictsUnaccusative Fragments.French.Predicates.motionCosSubjectProfile

theorem MartinSchaeferKastner2025.opposite_preferences_falsify_uniform_semantics : predictSePreference ControlLevel.limitedControl Animacy.human ≠ predictSePreference ControlLevel.inControl Animacy.human

The causation claim (@cite{labelle-1992}, @cite{labelle-doron-2010}) predicts that the presence vs. absence of se correlates with external vs. internal causation ACROSS THE BOARD. But the preferences documented here arise ONLY with ±se verbs (where the speaker has a choice) and ONLY with human DPs (where agent bias is active). This is incompatible with a semantic distinction between ±se forms.

Formally: the same verb class (±se) shows OPPOSITE preferences depending on control level — limited-control prefers −se, in-control prefers +se. A semantic account predicting uniform behavior for all ±se verbs is falsified.

structure MartinSchaeferKastner2025.ConditionMean : Type

Raw mean acceptability ratings (7-point Likert scale). Encoded as rationals (×1000) for decidable comparison. From experiments 1a and 1b.

• context : String
• animacy : Animacy
• plusSeMean : ℕ

  +se mean × 1000 (rational encoding)

• minusSeMean : ℕ

  −se mean × 1000 (rational encoding)

@[implicit_reducible]

instance MartinSchaeferKastner2025.instReprConditionMean : Repr ConditionMean

Equations

• MartinSchaeferKastner2025.instReprConditionMean = { reprPrec := MartinSchaeferKastner2025.instReprConditionMean.repr }

def MartinSchaeferKastner2025.instReprConditionMean.repr : ConditionMean → ℕ → Std.Format

Equations

• One or more equations did not get rendered due to their size.

@[implicit_reducible]

instance MartinSchaeferKastner2025.instDecidableEqConditionMean : DecidableEq ConditionMean

Equations

• MartinSchaeferKastner2025.instDecidableEqConditionMean = MartinSchaeferKastner2025.instDecidableEqConditionMean.decEq

def MartinSchaeferKastner2025.instDecidableEqConditionMean.decEq (x✝ x✝¹ : ConditionMean) : Decidable (x✝ = x✝¹)

Equations

• One or more equations did not get rendered due to their size.

def MartinSchaeferKastner2025.experiment1a_data : List ConditionMean

Experiment 1a: limited-control verbs (Table 2). Values are raw means × 1000.

Equations

• One or more equations did not get rendered due to their size.

def MartinSchaeferKastner2025.experiment1b_data : List ConditionMean

Experiment 1b: in-control verbs (Table 4). Values are raw means × 1000.

Equations

• One or more equations did not get rendered due to their size.

theorem MartinSchaeferKastner2025.exp1a_neutral_human_confirms_G1 : 3218 < 6526

G1 confirmed: for limited-control verbs with human DPs in the neutral context, −se ratings (6.526) exceed +se ratings (3.218).

theorem MartinSchaeferKastner2025.exp1b_neutral_human_confirms_G2 : 3628 < 5904

G2 confirmed: for in-control verbs with human DPs in the neutral context, +se ratings (5.904) exceed −se ratings (3.628).

theorem MartinSchaeferKastner2025.exp1a_reflexive_human_reversal : 3551 < 4904

G1 reverses in reflexive context: with limited-control verbs, human DPs in the reflexive context prefer +se (4.904 > 3.551). This is expected: the reflexive context forces an agentive construal, which only +se can express.

theorem MartinSchaeferKastner2025.exp1b_reflexive_human_confirms_G2 : 2994 < 5891

G2 holds in reflexive context too: with in-control verbs, human DPs in the reflexive context strongly prefer +se (5.891 > 2.994).