📄 Gerundive Complements in English : A Constraint-Based Analysis

Summary

This paper examines how a constraint-based (HPSG-style) grammar should handle English gerund phrases that serve as verb complements, as in I hate Pat's/Pat smoking cigars. Earlier generative accounts relied on empty categories, functional heads or derivations that split words in syntax, which violates lexical integrity. Building on Malouf (2000), Park adopts a lexical category gerund that sits under both noun and relational (the supertype shared with verb and adjective) in a multiple-inheritance hierarchy. This lets a gerund phrase take verbal complements and adverbs internally while appearing wherever NPs appear, without any category switch inside the tree. Using construction types and type constraints, the paper covers three subtypes of gerundive complements: POSS-ing, ACC-ing and PRO-ing. It also separates gerunds consistently from VP[VFORM participle]. Finally, it weighs a simpler alternative that treats gerunds as verbs carrying [VFORM gerund]. That alternative is cheaper but leaves their NP-like distribution unexplained. The work shows that "mixed" categories can be stated precisely in a formal grammar.

Key points

• Following Malouf (2000), gerund is a subtype of noun and of relational. It is not a subtype of verb, so gerund phrases (GP) appear wherever NPs do, but only in some VP positions.
• The genitive subject (POSS-ing) comes from a noun-possessive-construction. The accusative subject (ACC-ing) comes from a nonfinite-head-subject-construction, which is also a subtype of clause. The clause type requires its SUBJ to be empty, PRO or a gap.
• Verbs such as dislike, hate, mind, remember, discuss accept both ACC-ing and POSS-ing. Passivization (His smoking was hated by everyone) and pseudo-clefts show the GP is a direct-object NP.
• Malouf derives gerunds from present participles rather than base forms. This explains why non-existent forms like must-ing, can-ing or is-ing never occur, because finite forms cannot feed the rule.
• The VFORM-only alternative avoids adding a new part of speech. However, it makes the nominal distribution of gerunds an accident rather than a consequence. Choosing among lexical rules and handling the PRO subject are left for further study.

Interesting ideas

• I1. Model a two-faced word class as a type with two parents (nominal outside, verbal inside) instead of forcing one label or switching category mid-tree.
• I2. The case of the gerund's subject (genitive vs accusative) selects a different construction, yet both phrases name the same embedded event with the same agent.
• I3. Use syntactic diagnostics such as passivization and pseudo-clefting to check that a clausal-looking complement really behaves as an object NP.
• I4. Deriving one form from another, rather than from the base, can block impossible forms without extra stipulations.
• I5. Judge competing analyses by whether a pattern follows necessarily or only by accident, not just by surface coverage.

For Mova

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

• I1 (tokens): A VerbForm=Ger token could carry both a nominal-distribution flag and a verbal-valency flag. Later stages could then query whichever face they need.
• I2 (frames): When building frames, map both Pat's smoking and Pat smoking to an embedded event whose agent is Pat. A missing subject (PRO-ing) should resolve to the controller, e.g. the matrix subject.
• I3 (learn): Such transformations could serve as consistency checks on silver parses. A gerund labelled obj should survive passivization as nsubj:pass, and disagreement flags a likely error.
• I4 (tokens): In the morphology crate, generate -ing gerunds via the participle path. This way modals and finite forms never produce candidate lemmas like must-ing.
• I5 (world): When designing ontology and verb-class types, prefer inheritance that makes distributional facts fall out automatically. This keeps explanations derivable rather than listed case by case.