Linking Theory Predictions → Hand-Annotated θ-Roles #
@cite{kratzer-1996} @cite{rappaport-hovav-levin-1998}
Two accounts of argument realization make predictions about external
argument theta roles. This bridge file tests each account's predictions
independently against hand-annotated subjectTheta in the English
verb fragment. Both are instantiated as LinkingTheory (see
Theories/Interfaces/SyntaxSemantics/Linking.lean).
Account 1: Severing #
Voice flavor determines the theta role: Voice_AG → agent, Voice_CAUSE →
stimulus. The prediction chain is: verb → Voice selection → theta role.
Ctx = VoiceFlavor; predict ignores the verb.
The current Voice typology has only two θ-assigning flavors, so it can only distinguish agent from stimulus. It correctly predicts ~71% of verbs (all agents + all stimuli), but systematically fails for:
- Experiencer subjects (know, believe, enjoy,...): Voice_AG predicts agent, but the actual role is experiencer. The typology lacks a Voice_EXPERIENCER flavor.
- Theme subjects (arrive, glow, whisper,...): Voice_nonThematic predicts none (no external argument), but the subject IS the internal argument that moved. The severing account correctly predicts no external argument; the subject's theta role comes from V, not Voice.
Account 2: Lexicalist (@cite{levin-1993}, Rappaport @cite{rappaport-hovav-levin-1998}) #
The verb's lexical semantics determines the theta role, bypassing Voice.
Ctx = Unit; predict ignores structure. Uses attitudeBuilder,
causalSource, objectTheta, factivePresup, levinClass, unaccusative,
controlType — all verb-internal semantic properties.
Correctly predicts ~93% of verbs. Fails for 6 genuinely irreducible cases (sweep, remember, forget, dare, bother, hesitate) and 6 verbs with missing annotations.
The causative alternation is the structural prediction of Voice
severing: transitive "John broke the vase" has agentive Voice with an
agent in Spec,VoiceP; anticausative "The vase broke" has non-thematic
Voice with no specifier. Both share the same VP core. Event-structure
predictions are verified in Core/Voice.lean via buildDecomposition.
Equations
Instances For
Equations
Instances For
Equations
Instances For
Transitive: "John broke the vase"
[VoiceP John [Voice' Voice_AG [vP v [VP broke [DP the vase]]]]]
Equations
- One or more equations did not get rendered due to their size.
Instances For
Anticausative: "The vase broke"
[VoiceP Voice_∅ [vP v [VP broke [DP the vase]]]]
Equations
- One or more equations did not get rendered due to their size.
Instances For
Unaccusative: "The ship sank"
[VoiceP Voice_∅ [vP v [VP sank [DP the ship]]]]
Equations
- One or more equations did not get rendered due to their size.
Instances For
Middle: "The door opened"
[VoiceP Voice_MID [vP v [VP opened [DP the door]]]]
Equations
- One or more equations did not get rendered due to their size.
Instances For
Agent c-commands theme in the transitive.
Theme does NOT c-command agent.
Anticausative contains theme but no agent DP.
Unaccusative contains theme.
Middle contains theme.
The transitive has an agent DP; the anticausative does not.
Voice determines the alternation: agentive assigns θ, non-thematic does not. This is @cite{kratzer-1996}'s severing verified structurally on the tree derivations.