📄 Predicate-Argument Structure of English Adjectives

Summary

The paper asks how many arguments an English predicate adjective really has. Prepositional phrases and to-infinitives such as good at tennis, wise of Tom to go, easy for John to read and similar to mine make adjectives look two- or three-place. Ikeya extends his earlier framework (Ikeya [4]) in which degree adjectives are interpreted along three contextual dimensions. He then works through the adjective classes of Silva and Thompson: tough predicates, Class A (Types I and II), Class S, fond of adjectives and similar to adjectives. For each, he assigns predicate-argument structures built from constructed example sentences. His central device is the predicate modifier, an expression that narrows the predicate's denotation to a subset. Syntactically, in HPSG terms, these phrases are adjuncts; only the surface NP is subcategorized. The conclusion is that adjectives are basically one-place predicates. Their single argument can be an entity, a nominalized property or a proposition. This matters because it gives a principled test for telling genuine arguments from apparent ones.

Key points

• Degree adjectives are interpreted along three dimensions:
  • THEMATIC DIMENSION (TD, the respect: at basketball), taken from Bartsch;
  • COMPARATIVE DIMENSION (CD, the comparison class: for a short Japanese);
  • DEGREE DIMENSION (DD, the degree: very).
• In the unmarked case, only an implicit CD is present.
• Against Jacobson's three-place analysis of tough predicates, for NP and to VP are treated as predicate modifiers of type <<e,t>,<e,t>>. The adjective plus its modifiers forms a complex predicate.
• A single argument slot can hold one of three kinds of filler:
  • an entity (This book is easy to read);
  • a nominalized property (To read this book is tough);
  • a proposition (a complete for NP to VP clause as subject, as in the example where the wife's accepting the view is what would be tough).
• Thematic role (agent, experiencer, beneficiary, following Silva and Thompson) is kept separate from semantic function (predicate modifier). of NP and for NP have a thematic role but function as modifiers.
• Obligatory complements (fond of, worthy of) are also analysed as modifiers. similar to, parallel to and same as are treated as one-place predicates whose subject must denote two entities (These cars are the same).

Interesting ideas

• I1. Degree adjectives get three slots (respect, comparison class, degree), and a sentence's truth depends on filling them, explicitly or from context.
• I2. A predicate modifier restricts the predicate to a subset. By Berman's point, Mary is unpleasant to argue with does not mean Mary is unpleasant overall.
• I3. Separating a phrase's thematic role from its semantic function lets an of NP be the agent while still structurally modifying the adjective.
• I4. Extraposed it sentences, infinitival subjects and plain subjects map onto one structure that differs only in the argument kind (entity, nominalized property, proposition).
• I5. Relational adjectives like similar to are recast as one-place predicates over a plural subject.

For Mova

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

• I1 (world): Store adjective states with optional respect, comparison-class and degree fields, so that tall for his age is never recorded as plain tall.
• I2 (frames): Attach to VP and for NP to the adjective frame as restrictors, and block the inference to the unrestricted property, so contrasts like nice, but unpleasant to argue with stay consistent.
• I3 (frames): Keep two annotation layers in adjective frames, one for role labels and one for structural function, so explanations can say both who acted and what narrowed the predicate.
• I4 (frames): Normalize the expletive-it + to VP/that clause UD patterns and their fronted variants into one frame with a typed argument, so paraphrases yield identical meaning graphs.
• I5 (world): Store X is similar to Y and X and Y are similar as one symmetric fact over the group {X, Y}, so a QA question in either form finds it.