📄 Reflexivity in English: an STAG analysis

Summary

The paper models English reflexive pronouns (himself, herself) in Synchronous Tree Adjoining Grammar (STAG), where a syntactic derivation is paired with an isomorphic semantic one. Each reflexive is a two-part multi-component set. The first part substitutes into the argument slot. The second is a degenerate auxiliary tree that adjoins higher up and enforces φ-feature (person/number/gender) agreement with the antecedent, using Top/Bottom feature unification. In the semantics, following Reinhart and Reuland, the reflexive is not a link between two nominals. It is a λ-function that takes a predicate and merges two of its argument variables, which reduces the predicate's valence. Three variants (T′, V′, TP) cover reflexives bound by a subject, by an object, or across a clause boundary. The analysis handles mono-clausal sentences, raising with seems to, and ECM with believes. The main theoretical claim is that Chomsky's Condition A need not be stipulated. Its locality part follows from the formalism, and its c-command part follows from semantic type well-formedness.

Key points

• Each reflexive is a pair of trees, an approach adopted from Kallmeyer and Romero (2007) and, earlier, Ryant and Scheffler (2006):
  • an initial tree in the argument slot;
  • a degenerate auxiliary tree adjoining at T′, V′ or TP to check agreement.
• Semantic types differ by variant:
  • the T′-form has type <<e,<e,t>>,<e,t>>;
  • the V′-form has type <<e,<e,<e,t>>>,<e,<e,t>>>.
• The analysis uses unreduced λ-expressions on the semantic derived tree, following Han (2007), with λ-conversion applied at the end.
• Ill-formed sentences are blocked in two ways:
  • Jim introduces herself to Gillian fails because of a [3sgM]/[3sgF] feature clash;
  • using the wrong reflexive variant fails because functional application cannot proceed in the semantics.
• Raising (Julian seems to Miles to love himself) and ECM (Julian believes himself to be intelligent) are handled as follows:
  • both use multiple adjunction (Schabes and Shieber 1994);
  • ECM needs no special features on the verb, unlike Kallmeyer and Romero;
  • picture-NP reflexives are treated as logophors and set aside, and non-argument reflexives are left for future work.

Interesting ideas

• I1. A reflexive can be treated as an operator on the predicate that identifies two of its argument slots, rather than as a pointer to an antecedent.
• I2. Agreement checking is placed in a separate, content-free component, so the agreement constraint is kept apart from the meaning contribution.
• I3. Binding constraints are not hand-written; they fall out of type mismatches, so a derivation that cannot compose is simply ruled out.
• I4. Choosing among a small set of reflexive variants by the position of the antecedent (subject, object, higher clause) keeps binding local and enumerable.
• I5. The analysis tests itself with starred sentences that must be rejected, including agreement mismatches and finite-clause himself.

For Mova

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

• I1 (frames): When filling VerbNet roles, himself could be resolved by unifying the role slots of the same verb frame (e.g. Agent = Recipient), not by running general coreference.
• I2 (frames): The coreference module could check gender/number compatibility as an independent hard filter before scoring candidates, so that every rejection gets a clear, explainable reason.
• I3 (world): When building the text world, a candidate reading whose roles fail to compose type-wise can be discarded automatically, and the failure itself serves as the explanation.
• I4 (tree): After UD parsing, the dependency path from -self to the nsubj/obj of the same or the governing verb (xcomp for raising/ECM) could select among three explicit binding patterns.
• I5 (learn): Minimal ungrammatical or mis-bound reflexive pairs make good negative controls when evaluating the coreference and frame stages.