Psych Verb IC Bias — Proto-Role Bridge #
@cite{solstad-bott-2022} @cite{solstad-bott-2024} @cite{dowty-1991} @cite{kehler-2002}
Connects IC bias predictions to @cite{dowty-1991} proto-role infrastructure and coherence relations.
Core argument #
IC bias tracks the stimulus argument: explanations in because-continuations target the participant whose entailment profile includes causation (the stimulus/cause), regardless of whether that participant is the subject or the object.
- StimExp (B&R Class II): stimulus = subject → NP1 bias (87.4%, Table 1)
- ExpStim (B&R Class I): stimulus = object → NP2 bias (96.0%, Table 1)
Entailment profiles by verb class #
| Class | Subject profile | P-Agent entailments |
|---|---|---|
| StimExp | C + IE (stimulus/causer) | causation, indep.exist. |
| ExpStim | S + IE (experiencer) | sentience, indep.exist. |
| AgPat | V + S + C + M + IE | all five |
Stimulus-experiencer verb subject profile: causation + independent existence. The subject is a stimulus/cause (B&R Class II, Levin 31.1 amuse class). @cite{solstad-bott-2022}: STIM-EXP verbs show NP1 I-Caus bias.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Stimulus-experiencer verb object profile: sentience + independent existence. The object is an experiencer.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Experiencer-stimulus verb subject profile: sentience + independent existence. The subject is an experiencer (B&R Class I, temere class). @cite{solstad-bott-2022}: EXP-STIM verbs show NP2 I-Caus bias.
Equations
- One or more equations did not get rendered due to their size.
Instances For
Experiencer-stimulus verb object profile: causation + independent existence. The object is a stimulus (cause of the experience).
Equations
- One or more equations did not get rendered due to their size.
Instances For
Agent-patient verb subject profile: full agent (all 5 P-Agent). Identical to existing kickSubjectProfile.
Equations
- One or more equations did not get rendered due to their size.
Instances For
StimExp subject profile = ExpStim object profile (both are stimulus/C+IE).
ExpStim subject profile = StimExp object profile (both are experiencer/S+IE).
This is the B&R theta-role reversal expressed at the proto-role level: Class I and Class II swap the same two profiles between subject and object.
StimExp subjects pass the do-test (they have causation). "What the noise did was frighten John" is grammatical because the subject has the causation entailment (Dowty's P-Agent (c)).
ExpStim subjects fail the do-test (experiencers lack volition, causation, and movement). "??What Mary did was admire John" is marginal.
AgPat subjects pass the do-test (full agents).
StimExp subject profile matches ThetaRole.stimulus's canonical profile.
ExpStim subject profile matches ThetaRole.experiencer's canonical profile.
AgPat subject profile matches ThetaRole.agent's canonical profile.
The Explanation relation (triggered by "because") selects for causes.
IC bias prediction: under Explanation (because), the continuation targets the STIMULUS argument — the participant whose entailment profile includes causation.
- StimExp: subject has causation → explanation about subject → NP1
- ExpStim: subject has sentience only → explanation about object → NP2
- AgPat: subject has causation (+ volition, etc.) → NP1
Equations
- One or more equations did not get rendered due to their size.
Instances For
StimExp predicted as NP1 (stimulus subject has causation).
ExpStim predicted as NP2 (experiencer subject lacks causation).
AgPat predicted as NP1 (agent subject — default).
The prediction matches the empirical data for all tested classes.
The IC reversal (StimExp→NP1, ExpStim→NP2) and the transfer verb goal bias are both instances of the same deeper pattern: swapping which argument carries a discourse-prominent thematic role reverses the discourse bias direction.
For IC: swapping stimulus between subject (StimExp) and object (ExpStim) reverses the IC bias from NP1 to NP2. For transfer: swapping goal between subject and nonsubject doesn't eliminate the goal bias — goals still get more pronouns in BOTH positions.
The IC reversal is the stronger demonstration: it shows the bias direction is ENTIRELY determined by the thematic role, not the grammatical position. @cite{rosa-arnold-2017}'s data corroborates this by showing that thematic role affects form even when grammatical role is held constant, violating @cite{kehler-rohde-2013}'s independence hypothesis.
Coherence relations select for COMPLEMENTARY thematic roles in the two phenomena, demonstrating that the coherence–role interaction is systematic rather than accidental:
Explanation (because) → selects CAUSE → stimulus in psych verbs Occasion/Result → selects ENDPOINT → goal in transfer verbs
This complementarity is predicted by the semantics of the coherence relations: Explanation asks "why did this happen?" (→ cause), while Occasion/Result asks "what happened next?" (→ endpoint). The same verb may participate in both patterns depending on which coherence relation the continuation establishes.
The empirical signatures differ accordingly:
- @cite{solstad-bott-2024}: because-continuations target the stimulus
- @cite{rosa-arnold-2017}: Occasion/Result amplifies goal bias while Other (including Explanation) does not
@cite{kehler-rohde-2013}'s Table 2 establishes that Explanation coherence relations are Source-biased (80% Source for transfer verbs). This study's IC data instantiates the same mechanism for psych verbs: Explanation (triggered by "because") selects for causes, and IC bias tracks whichever argument carries the causation entailment — the stimulus.
K&R's coherence-marginalized model (eq. 9) predicts that IC bias should be driven by P(referent | Explanation), not by grammatical position. The StimExp/ExpStim reversal confirms this: the bias direction is ENTIRELY determined by which argument is the stimulus, regardless of subject/object position.
@cite{kehler-rohde-2013}'s key structural claim is that coherence relations and referential form contribute to DIFFERENT terms in Bayes' rule: P(referent) vs P(pronoun|referent). The IC data provides the strongest evidence for the P(referent) side:
- K&R: P(referent) = Σ_CR P(CR) × P(referent | CR)
- IC context: "because" sets P(Explanation) ≈ 1
- Therefore: P(referent) ≈ P(referent | Explanation)
- P(referent | Explanation) = whichever argument has causation
This is exactly what the IC data shows: under "because" prompts, the bias aligns perfectly with the causation entailment profile, as predicted by a model where P(CR=Explanation) dominates the mixture and P(referent|Explanation) tracks the stimulus.