📄 Binding Variables in English: An Analysis Using Delayed Tree Locality

Summary

The paper models English pronouns that act as bound variables, such as a possessive pronoun bound by a quantified subject like "every girl." It works in Synchronous TAG (STAG), which pairs syntax trees with semantics trees. The difficulty is that the pronoun and its quantifier binder can sit several predicates apart, while standard TAG composes only locally. The authors treat each bound pronoun as a multi-component (MC) set in both syntax and semantics. One piece fills the argument slot. A defective auxiliary piece attaches to the binder, where agreement features are unified and the variable is re-bound. The pieces compose through delayed tree-local MC-TAG. For core cases, at most two simultaneous delays are needed, which keeps the grammar weakly equivalent to flexible composition. Three constraints filter out bad readings:

• a semantic ban on unbound variables,
• a syntactic c-command requirement between the two pronoun pieces, which handles weak crossover,
• an anti-locality bound on delay size.

Sentences with several bound pronouns inside one DP need more than two simultaneous delays. The authors argue this favours delayed locality over flexible composition.

Key points

• The framework is STAG (Shieber 1994) with feature agreement, using Delayed Tree-Local MC-TAG (Chiang and Scheffler 2008). Quantifiers get a generalised-quantifier semantics split into a variable tree and a scope tree.
• A delay is the set of derivation-tree nodes on the shortest path between the members of an MC set, excluding their lowest common dominating node.
• Anti-locality constraint: the delay of a bound-variable MC set must contain at least four nodes. This rules out a pronoun bound by its own clause-mate subject, where a reflexive is required, and also strong crossover. Both of those configurations give a delay of three.
• Weak crossover (a pronoun inside the subject's possessor, bound by a quantified object) still passes the size check, with a delay of four. It is blocked instead by requiring the defective DP* piece to c-command the argument DP piece.
• A DP containing two or three bound pronouns pushes one derivation node into 3 or 4 simultaneous delays. So no fixed k suffices in principle. The authors suggest the number of delays may track processing difficulty.

Interesting ideas

• I1. A bound pronoun is split into two linked parts: one occupies the argument position, the other attaches at the binder and merges agreement features there. Binding thus becomes an explicit, checkable link rather than a guess.
• I2. Locality is measured as the size of the derivation path between a pronoun and its antecedent, and the path must exceed a minimum. This yields a simple numeric test for when a plain pronoun cannot pick up a nearby antecedent.
• I3. Some quantifier scope orders leave a variable free, and these are discarded only after composition. The grammar can overgenerate while a cheap semantic filter removes the bad orders.
• I4. Every constraint is justified by at least one ungrammatical sentence that only it blocks, even though some bad sentences violate several.
• I5. The count of simultaneous delays is proposed as a proxy for how hard a sentence is to interpret.

For Mova

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

• I1 (frames): coreference could store a quantified antecedent as a separate binder node, with each pronoun mention linked to it. Gender and number unification at that link would explain why the link was chosen or rejected.
• I2 (frames): a UD-path-length or co-argument check could forbid non-reflexive pronouns from resolving to a subject of the same predicate, and the check could be reported in the explanation.
• I3 (world): when building the world of a text with several quantified entities, reject scope orderings that leave a pronoun's referent unanchored, instead of hand-coding the order.
• I4 (learn): each repair rule or filter could require its own negative control example that fails without it, so redundant rules are detected and removed.
• I5 (answer): the number of nested bound pronouns or simultaneous dependencies could serve as a difficulty score for comprehension test items, or flag answers that deserve low confidence.