@cite{rosa-arnold-2017}

@cite{kehler-rohde-2013} @cite{kehler-2002} @cite{arnold-wasow-losongco-ginstrom-2000}

Predictability Affects Production: Thematic Roles Can Affect Reference Form Selection. Journal of Memory and Language 94, 43–60.

Core Argument

Speakers use more pronouns for goals than sources of transfer verbs, across three experimental paradigms (event-retelling, sentence completion × 2). A rating study confirms that goal characters are more predictable next-mentions (71% chose the goal as likely next-mention; separately, only 54% chose the subject, suggesting subjecthood is a weaker predictor of next-mention than thematic role). This establishes that thematic roles affect referential form selection via predictability, contrary to claims that thematic roles do not affect form.

Key Findings

#	Finding	Status
1	Goals get more pronouns than sources (all 3 exps)	data
2	Subjects get more pronouns than nonsubjects (all 3 exps)	data
3	Goals are more predictable next-mentions (71% vs 54%)	data
4	Occasion/Result coherence amplifies goal bias (Exp 2)	data
5	Goal bias robust across paradigms	data
6	Transfer verbs assign Goal to indirect object	rfl
7	Occasion/Result is contiguity/causeEffect	rfl
8	Goal > Source mirrors IC next-mention mechanism	cross-study
9	Form reduction feeds into ordering (Arnold et al. 2000)	cross-study

Debate with @cite{kehler-rohde-2013}

Kehler & Rohde decompose pronoun interpretation via Bayes' rule:

P(referent | pronoun) ∝ P(pronoun | referent) × P(referent)

They propose two independent factors:

• P(referent): coherence-driven next-mention bias (sensitive to thematic roles)
• P(pronoun | referent): centering-driven form bias (sensitive to subjecthood ONLY)

This predicts thematic roles should NOT affect pronominalization rate. @cite{rosa-arnold-2017} directly challenges this independence: goals get more pronouns than sources even controlling for grammatical role, showing P(pronoun | referent) is also sensitive to thematic role predictability.

Connection to @cite{arnold-wasow-losongco-ginstrom-2000}

The same verb ("give"), the same construction (dative/transfer), but different dependent variables: Arnold et al. (2000) study position (heavy NP shift, dative alternation), Rosa & Arnold (2017) study form (pronoun vs name). Both are production choices along the same NP weight/reduction dimension.

inductive RosaArnold2017.TransferRole : Type

Thematic role of the referent in a transfer verb event.

• goal : TransferRole
• source : TransferRole

@[implicit_reducible]

instance RosaArnold2017.instDecidableEqTransferRole : DecidableEq TransferRole

Equations

• RosaArnold2017.instDecidableEqTransferRole x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

@[implicit_reducible]

instance RosaArnold2017.instReprTransferRole : Repr TransferRole

Equations

• RosaArnold2017.instReprTransferRole = { reprPrec := RosaArnold2017.instReprTransferRole.repr }

def RosaArnold2017.instReprTransferRole.repr : TransferRole → ℕ → Std.Format

Equations

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

inductive RosaArnold2017.GramRole : Type

Grammatical role of the referent in the prior sentence.

• subject : GramRole
• nonsubject : GramRole

@[implicit_reducible]

instance RosaArnold2017.instDecidableEqGramRole : DecidableEq GramRole

Equations

• RosaArnold2017.instDecidableEqGramRole x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

def RosaArnold2017.instReprGramRole.repr : GramRole → ℕ → Std.Format

Equations

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

@[implicit_reducible]

instance RosaArnold2017.instReprGramRole : Repr GramRole

Equations

• RosaArnold2017.instReprGramRole = { reprPrec := RosaArnold2017.instReprGramRole.repr }

inductive RosaArnold2017.GenderContext : Type

Gender match between referents (affects ambiguity of pronouns).

• sameGender : GenderContext
• differentGender : GenderContext

@[implicit_reducible]

instance RosaArnold2017.instDecidableEqGenderContext : DecidableEq GenderContext

Equations

• RosaArnold2017.instDecidableEqGenderContext x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

def RosaArnold2017.instReprGenderContext.repr : GenderContext → ℕ → Std.Format

Equations

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

@[implicit_reducible]

instance RosaArnold2017.instReprGenderContext : Repr GenderContext

Equations

• RosaArnold2017.instReprGenderContext = { reprPrec := RosaArnold2017.instReprGenderContext.repr }

inductive RosaArnold2017.Paradigm : Type

Experimental paradigm.

• eventRetelling : Paradigm
• sentenceCompletion : Paradigm
• renamedCompletion : Paradigm

@[implicit_reducible]

instance RosaArnold2017.instDecidableEqParadigm : DecidableEq Paradigm

Equations

• RosaArnold2017.instDecidableEqParadigm x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

@[implicit_reducible]

instance RosaArnold2017.instReprParadigm : Repr Paradigm

Equations

• RosaArnold2017.instReprParadigm = { reprPrec := RosaArnold2017.instReprParadigm.repr }

def RosaArnold2017.instReprParadigm.repr : Paradigm → ℕ → Std.Format

Equations

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

structure RosaArnold2017.Condition : Type

Experimental condition: fully crossed design.

• role : TransferRole
• gram : GramRole
• gender : GenderContext

@[implicit_reducible]

instance RosaArnold2017.instDecidableEqCondition : DecidableEq Condition

Equations

• RosaArnold2017.instDecidableEqCondition = RosaArnold2017.instDecidableEqCondition.decEq

def RosaArnold2017.instDecidableEqCondition.decEq (x✝ x✝¹ : Condition) : Decidable (x✝ = x✝¹)

Equations

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

@[implicit_reducible]

instance RosaArnold2017.instReprCondition : Repr Condition

Equations

• RosaArnold2017.instReprCondition = { reprPrec := RosaArnold2017.instReprCondition.repr }

def RosaArnold2017.instReprCondition.repr : Condition → ℕ → Std.Format

Equations

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

structure RosaArnold2017.PronounRate : Type

Pronoun rate data point: percentage of pronoun use in a condition.

• paradigm : Paradigm
• cond : Condition
• percent : ℕ

def RosaArnold2017.instReprPronounRate.repr : PronounRate → ℕ → Std.Format

Equations

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

@[implicit_reducible]

instance RosaArnold2017.instReprPronounRate : Repr PronounRate

Equations

• RosaArnold2017.instReprPronounRate = { reprPrec := RosaArnold2017.instReprPronounRate.repr }

def RosaArnold2017.exp1_goal_subj_diff : PronounRate

Equations

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

def RosaArnold2017.exp1_source_subj_diff : PronounRate

Equations

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

def RosaArnold2017.exp1_goal_nonsub_diff : PronounRate

Equations

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

def RosaArnold2017.exp1_source_nonsub_diff : PronounRate

Equations

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

def RosaArnold2017.exp2_goal_subj_diff : PronounRate

Equations

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

def RosaArnold2017.exp2_source_subj_diff : PronounRate

Equations

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

def RosaArnold2017.exp2_goal_nonsub_diff : PronounRate

Equations

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

def RosaArnold2017.exp2_source_nonsub_diff : PronounRate

Equations

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

def RosaArnold2017.exp3_goal_subj_diff : PronounRate

Equations

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

def RosaArnold2017.exp3_source_subj_diff : PronounRate

Equations

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

def RosaArnold2017.exp3_goal_nonsub_diff : PronounRate

Equations

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

def RosaArnold2017.exp3_source_nonsub_diff : PronounRate

Equations

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

def RosaArnold2017.exp3_goal_nonsub_same : PronounRate

Equations

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

def RosaArnold2017.exp3_source_nonsub_same : PronounRate

Equations

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

structure RosaArnold2017.NextMentionRating : Type

Next-mention rating: percentage of participants choosing this role as the character most likely to be talked about next.

• role : TransferRole
• percent : ℕ

def RosaArnold2017.instReprNextMentionRating.repr : NextMentionRating → ℕ → Std.Format

Equations

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

@[implicit_reducible]

instance RosaArnold2017.instReprNextMentionRating : Repr NextMentionRating

Equations

• RosaArnold2017.instReprNextMentionRating = { reprPrec := RosaArnold2017.instReprNextMentionRating.repr }

def RosaArnold2017.nextMention_goal : NextMentionRating

71% of raters chose the goal character as most likely to be mentioned next (t(19)=4.91, p<.0001).

Equations

• RosaArnold2017.nextMention_goal = { role := RosaArnold2017.TransferRole.goal, percent := 71 }

def RosaArnold2017.nextMention_subject_pct : ℕ

54% of raters chose the subject (not significant, p>.1).

Equations

• RosaArnold2017.nextMention_subject_pct = 54

theorem RosaArnold2017.goals_more_predictable : nextMention_goal.percent > 100 - nextMention_goal.percent

Goals are more predictable next-mentions than sources.

theorem RosaArnold2017.goal_bias_stronger_than_subj_bias : nextMention_goal.percent > nextMention_subject_pct

Thematic role (goal: 71%) is a stronger next-mention predictor than grammatical role (subject: 54%). This supports the paper's core claim that predictability driven by thematic roles matters for production.

theorem RosaArnold2017.exp1_goal_gt_source_subj : exp1_goal_subj_diff.percent > exp1_source_subj_diff.percent

Goal > Source in pronoun rate: verified in every paradigm. Exp 1: 64 vs 37 (subject, different-gender). Exp 2: 55 vs 33 (nonsubject, different-gender). Exp 3: 33 vs 10 (nonsubject, same-gender — strongest interaction cell).

theorem RosaArnold2017.exp2_goal_gt_source_nonsub : exp2_goal_nonsub_diff.percent > exp2_source_nonsub_diff.percent

theorem RosaArnold2017.exp3_goal_gt_source_nonsub_same : exp3_goal_nonsub_same.percent > exp3_source_nonsub_same.percent

Exp 3 strongest cell: nonsubject same-gender shows 33% vs 10%.

theorem RosaArnold2017.exp1_subj_gt_nonsub : exp1_goal_subj_diff.percent > exp1_goal_nonsub_diff.percent

Subject > Nonsubject in pronoun rate (orthogonal to thematic role). Exp 1: 64 vs 31 for goals.

inductive RosaArnold2017.CoherenceCategory : Type

Coherence relation categories used in Exp 2 coding.

• occasionResult : CoherenceCategory
• other : CoherenceCategory

@[implicit_reducible]

instance RosaArnold2017.instDecidableEqCoherenceCategory : DecidableEq CoherenceCategory

Equations

• RosaArnold2017.instDecidableEqCoherenceCategory x✝ y✝ = if h : x✝.ctorIdx = y✝.ctorIdx then isTrue ⋯ else isFalse ⋯

@[implicit_reducible]

instance RosaArnold2017.instReprCoherenceCategory : Repr CoherenceCategory

Equations

• RosaArnold2017.instReprCoherenceCategory = { reprPrec := RosaArnold2017.instReprCoherenceCategory.repr }

def RosaArnold2017.instReprCoherenceCategory.repr : CoherenceCategory → ℕ → Std.Format

Equations

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

structure RosaArnold2017.CoherenceInteraction : Type

Exp 2 coherence interaction: Goal vs Source effect by coherence category. Occasion/Result: β=1.22 (0.40), t=3.06, p=.002 — significant. Other: β=0.86 (0.55), t=1.56, p=.12 — not significant.

• category : CoherenceCategory
• goalSourceBeta : ℤ
• significant : Bool

@[implicit_reducible]

instance RosaArnold2017.instReprCoherenceInteraction : Repr CoherenceInteraction

Equations

• RosaArnold2017.instReprCoherenceInteraction = { reprPrec := RosaArnold2017.instReprCoherenceInteraction.repr }

def RosaArnold2017.instReprCoherenceInteraction.repr : CoherenceInteraction → ℕ → Std.Format

Equations

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

def RosaArnold2017.occasionResult_interaction : CoherenceInteraction

Equations

• RosaArnold2017.occasionResult_interaction = { category := RosaArnold2017.CoherenceCategory.occasionResult, goalSourceBeta := 122, significant := true }

def RosaArnold2017.other_interaction : CoherenceInteraction

Equations

• RosaArnold2017.other_interaction = { category := RosaArnold2017.CoherenceCategory.other, goalSourceBeta := 86, significant := false }

theorem RosaArnold2017.coherence_amplifies_goal_bias : occasionResult_interaction.goalSourceBeta > other_interaction.goalSourceBeta

The goal bias is larger for Occasion/Result than Other coherence.

def RosaArnold2017.kehlerRohdeIndependence (pronounGivenReferent : TransferRole → GramRole → ℕ) (gram : GramRole) : Prop

@cite{kehler-rohde-2013}'s independence hypothesis: P(pronoun | referent) depends only on grammatical/topichood status, NOT on thematic role or coherence-driven predictability.

This predicts that pronominalization rate should be constant across thematic roles when grammatical role is held constant.

Equations

• RosaArnold2017.kehlerRohdeIndependence pronounGivenReferent gram = (pronounGivenReferent RosaArnold2017.TransferRole.goal gram = pronounGivenReferent RosaArnold2017.TransferRole.source gram)

theorem RosaArnold2017.independence_violated_exp1_subj : ¬kehlerRohdeIndependence (fun (role : TransferRole) (gram : GramRole) => match role, gram with | TransferRole.goal, GramRole.subject => 64 | TransferRole.source, GramRole.subject => 37 | TransferRole.goal, GramRole.nonsubject => 31 | TransferRole.source, GramRole.nonsubject => 18) GramRole.subject

@cite{rosa-arnold-2017}'s challenge: goals get more pronouns than sources even in the same grammatical position. This violates the independence hypothesis. Verified in Exp 1 subject condition.

def RosaArnold2017.transferNextMention : ThetaRole → Features.NextMentionBias

Transfer verb next-mention prediction: Goal arguments have higher next-mention bias than Source arguments in narrative (Occasion/Result) continuations. This maps the ThetaRole distinction to NextMentionBias.

Equations

• RosaArnold2017.transferNextMention ThetaRole.goal = Features.NextMentionBias.high
• RosaArnold2017.transferNextMention ThetaRole.source = Features.NextMentionBias.low
• RosaArnold2017.transferNextMention x✝ = Features.NextMentionBias.low

theorem RosaArnold2017.goal_predicts_pronoun : (transferNextMention ThetaRole.goal).predictedForm = Features.AccessibilityLevel.unstressedPron

Goal → pronoun, Source → name: the predicted referential form for transfer verb arguments follows from next-mention bias.

theorem RosaArnold2017.source_predicts_name : (transferNextMention ThetaRole.source).predictedForm = Features.AccessibilityLevel.fullName

theorem RosaArnold2017.occasion_is_contiguity : Core.Discourse.Coherence.CoherenceRelation.occasion.toClass = Core.Discourse.Coherence.CoherenceClass.contiguity

Occasion relations focus on the end state of the previous event. For transfer verbs, the Goal is the endpoint — the entity in the final state after transfer. This is why Occasion/Result coherence amplifies the Goal bias.

Occasion is a contiguity relation; Result is cause–effect. Both focus on what happens AFTER the event, favoring the Goal.

theorem RosaArnold2017.result_is_cause_effect : Core.Discourse.Coherence.CoherenceRelation.result.toClass = Core.Discourse.Coherence.CoherenceClass.causeEffect

theorem RosaArnold2017.pronoun_more_reduced : Features.AccessibilityLevel.unstressedPron.rank > Features.AccessibilityLevel.fullName.rank

The same transfer verb "give" is studied for both referential form (@cite{rosa-arnold-2017}) and constituent ordering (@cite{arnold-wasow-losongco-ginstrom-2000}). Pronouns are more reduced than names on the accessibility scale, and at most as heavy by word count. The referential form choice connects to ordering:

Rosa & Arnold: Goal → pronoun (reduced) Arnold et al. 2000: light/reduced NP → early position

Together: Goal → pronoun → early position. The referential form choice mediates between thematic role and syntactic position.

theorem RosaArnold2017.pronoun_at_most_as_heavy : Features.ReferentialForm.typicalWeight Features.AccessibilityLevel.unstressedPron ≤ Features.ReferentialForm.typicalWeight Features.AccessibilityLevel.fullName

theorem RosaArnold2017.goal_more_reduced_than_source : Features.AccessibilityLevel.rank (transferNextMention ThetaRole.goal).predictedForm > Features.AccessibilityLevel.rank (transferNextMention ThetaRole.source).predictedForm

The goal argument receives a MORE REDUCED referential form than the source argument. This derived contrast — not the individual predictions — is the empirical content of @cite{rosa-arnold-2017}.

The same reduction asymmetry creates a weight asymmetry: goal arguments surface as lighter NPs (pronouns) while source arguments surface as heavier NPs (names). @cite{arnold-wasow-losongco-ginstrom-2000} prove that exactly this weight dimension independently predicts constituent ordering in dative constructions with the same verb. So thematic roles affect ordering through referential form reduction.

theorem RosaArnold2017.dual_path_to_ordering : (transferNextMention ThetaRole.goal).predictedForm ≠ (transferNextMention ThetaRole.source).predictedForm ∧ nextMention_goal.percent > 100 - nextMention_goal.percent ∧ Core.Constraint.harmonyDominates (ArnoldEtAl2000.grammar 1 0) (ArnoldEtAl2000.heavyGoalContrast, ArnoldEtAl2000.Order.goalLast) (ArnoldEtAl2000.heavyGoalContrast, ArnoldEtAl2000.Order.themeLast) ∧ Core.Constraint.harmonyDominates (ArnoldEtAl2000.grammar 0 1) (ArnoldEtAl2000.newThemeContrast, ArnoldEtAl2000.Order.themeLast) (ArnoldEtAl2000.newThemeContrast, ArnoldEtAl2000.Order.goalLast)

@cite{arnold-wasow-losongco-ginstrom-2000} show that heaviness and newness BOTH independently predict ordering. @cite{rosa-arnold-2017} show thematic roles affect BOTH form (the heaviness dimension) and predictability (the newness dimension). Thematic roles therefore have a dual path to constituent ordering:

Path 1 (via form): θ-role → form reduction → lighter NP → earlier Path 2 (via predictability): θ-role → next-mention bias → given-like → earlier

This theorem derives the existence of both paths: the goal/source contrast produces different predicted forms (Path 1 input), and goals are more predictable than sources (Path 2 input). Arnold et al. confirm that both receiving dimensions independently matter — bridged via the MaxEnt independence theorems heaviness_independently_predicts and newness_independently_predicts, instantiated at the heaviness-only and newness-only stimulus contrasts.

theorem RosaArnold2017.independence_violated_bridges_to_KR : KehlerRohde2013.pron_passive_subj > KehlerRohde2013.pron_active_subj ∧ exp1_goal_subj_diff.percent > exp1_source_subj_diff.percent

@cite{kehler-rohde-2013}'s Bayesian decomposition predicts that P(pronoun | referent) depends only on topichood, not on semantic factors like thematic role. This study directly violates that prediction: goals get more pronouns than sources in the same grammatical position (Exp 1: 64% vs 37% for subjects).

The violation connects to K&R's Table 9 data, which shows that P(pronoun | referent) DOES vary with topichood (passive subject 87% vs active subject 62%). Rosa & Arnold extend this: thematic roles also contribute to topichood/predictability, not just syntactic construction.

Substrate-level explanation: KehlerRohde2013.cb_topichood_dissociation_under_voice (§12 of Phenomena/Reference/Studies/KehlerRohde2013.lean) exhibits the structural reason: under Kameyama's GR ranker, Centering's cb is voice-blind, so any pronominalization gradient between active and passive subjects must be carried by a signal external to cb — exactly the gap Rosa & Arnold's experiment measures, and exactly the opening in K&R's Independence Hypothesis their data exploits.

theorem RosaArnold2017.coherence_interaction_matches_KR_biases : KehlerRohde2013.cr_occasion.sourceGivenCR < 50 ∧ KehlerRohde2013.cr_explanation.sourceGivenCR > 50 ∧ occasionResult_interaction.significant = true ∧ other_interaction.significant = false

K&R's Table 2 shows that Occasion and Result are Goal-biased (18% and 8% Source respectively). This study's Exp 2 coherence interaction confirms: Occasion/Result continuations amplify the goal bias (β=1.22, p=.002), while Other coherence (including Explanation, which K&R show is Source-biased at 80%) does not reach significance (β=0.86, p=.12). The coherence-specific biases from K&R's passage completion data predict the interaction pattern in this study's sentence completion data.