📄 Reconsidering Raising and Experiencers in English

Summary

The paper looks at how Tree Adjoining Grammar (TAG) should handle the raising verb seem when it carries an experiencer phrase such as to me. That phrase can sit right after seem (medial) or at the end of the clause (final). Frank (2002) proposed a flat, three-branch auxiliary tree for this. Storoshenko asks two questions: whether the experiencer is a true argument of seem, and whether a flat tree can explain how it behaves.

He first shows that the experiencer is an argument in both positions:

• Simple VP-adjunction cannot derive the medial word order.
• Extraction tests show the final position is not an island.

He then applies Larson-style c-command diagnostics. The medial experiencer c-commands the embedded object; the final one does not. A flat structure cannot explain this difference.

He therefore proposes a VP-shell analysis, borrowed from work on ditransitives. There are two lexically separate seem trees, and the verb moves up within its own elementary tree. Apparent word-order asymmetries with small clauses and finite complements are explained by phonological weight and ambiguity. This matters because it brings raising in TAG into line with modern ditransitive syntax, and it opens extra adjunction sites inside the seem tree.

Key points

• Data: hand-built English minimal pairs judged by native speakers. Examples include John seems to me to like coffee versus John seems to like coffee to me. There is no corpus or treebank.
• Argument tests for the final experiencer: relativisation, stripping, and a resumptive-pronoun repair of a complex-NP violation. All three show it can be extracted.
• c-command tests: NPI licensing, bound-variable readings, and Superiority. They succeed with the medial experiencer and fail with the final one.
• Proposal: two unrelated seem trees, each fixing the experiencer's position. This follows Harley (2002) rather than Larson's (1988) derivational link. A lower VP holds the experiencer and the T′ foot node, and seem raises to an upper VP that has no agent.
• The lower VP also gives a landing site for manner and temporal adjuncts that modify seem, possibly even when there is no experiencer. For instance, the paper's example inserts a duration clause about how long the speakers had known John between past-tense seemed and its complement to like coffee.

Interesting ideas

• I1. Decide whether a phrase is an argument or an adjunct with structural probes, such as extraction, NPI licensing and binding, rather than by intuition; this gives a reproducible criterion where labels are disputed.
• I2. With raising, the surface subject gets its semantic role only from the embedded predicate, while seem adds an experiencer who holds the impression.
• I3. A to-PP can mark a perspective holder even without seem, as in John likes coffee to me. This creates real attachment ambiguity in finite-clause cases.
• I4. Position choice can be driven by heaviness: speakers place the experiencer medially or finally depending on how heavy each competing constituent is.
• I5. Treat word-order variants as two separate lexical entries rather than one entry plus a transformation, which keeps each structure simple and local.

For Mova

How these ideas could serve Mova, our small open language model — its files, stages and tests:

• I1 (frames): Use the same probes to justify treating the to-PP of seem/appear as the Experiencer of the raising verb, not as a role of the embedded verb such as like.
• I2 (world): John seems to me to like coffee should store "John likes coffee" as a belief held by the experiencer, not as a fact about the world.
• I3 (tree): Add a repair or conversion rule that flags a clause-final to-PP after a seem complement as possibly attaching to either verb, and keep both parses for downstream disambiguation.
• I4 (back): When generating raising sentences with experiencers, compare constituent lengths and put the heavier one last, so the output sounds natural.
• I5 (frames): Store two frame templates for seem (medial and final experiencer) that map to the same roles, so frame matching never needs movement rules.