Polyfunctionality of what in English Transparent Free Relative Constructions

A Construction Grammar Approach

Jong-Bok Kim and Brett Reynolds

2026-04-08

How does what manage to be singular and plural?

  1. [What appears to be a painting] turns into something else. (singular)
  2. [What we call coincidences] are limited. (plural)

The predicative nucleus determines agreement, not what itself.

The polyfunctionality of what

Construction Example What as…
Interrogative What did you see? interrogative pronoun
Standard FR I’ll eat what you’re having. free relative pronoun (nominal)
Exclamative What a mess! exclamative
TFR She’s what they call ecstatic. underspecified; resolved by nucleus

In TFRs, what is genuinely underspecified. Its category is resolved by the predicative nucleus.

NB: only independent what (without a following noun) introduces TFRs. *What food they call delicious is out.

TFR data: distributional range

  1. You’re definitely not (what anyone would describe as) ecstatic. (AP)
  2. I begin to work (what I would call) creatively. (AdvP)
  3. She wasn’t (what she’d call) in love with Sam. (PP)
  4. You make him responsible for (what I call) trumping the center. (NP)

The nucleus determines the category of the entire TFR.

Coordination and agreement

Coordination (COCA)
  • Connelly is pretty excited about his life and (what he considers) the perfect job.
  • Omer Stewart counted reasons for prehistoric and (what he described as) primitive uses of fire.
Agreement
  • [What appears to be trousers] are / *is really leggings.
  • Plural nucleus -> plural verb.
  • Standard FRs do not show this flexibility.

Both diagnostics point to the external nucleus as head.

Prepositional selection and headedness

Selectional test
  • in / *at what linguists call a Northern dialect
  • The matrix predicate selects the preposition, not the embedded clause.

If the nucleus stayed inside the clause, this contrast would be much harder to derive.

Why the nucleus has to be outside the clause

SFR (what is head)

NP -> [NP what] [S/NP they call ___]

TFR (nucleus is head)

AP -> [S/AP what they call ___] [AP systematic]

Grosu (2003) derives transparency endocentrically via feature percolation. The exocentric analysis is motivated by idiom and extraction evidence.

Idiom split

Nick has made (what one may call) significant headway.

The idiomatic reading requires locality between the matrix verb and complement.

Verb inventory and coupled constraints

Group Verbs Non-NP nucleus?
Attributional call, consider, describe as, regard as, take to be yes: AP, AdvP, PP
Evidential seem, appear NP only (0 in COCA)

Category transparency is confined to the attributional verbs. Take to be has a copula but patterns with call. The copula is not the blocker.

The same lexical group that supports source attribution also supports the full category range. In SBCG, that coupling belongs in the construction.

Follow-up checks: the split survives extension

Group Predicate Current status
Attributional call, consider, describe as, regard as, take to be non-NP nuclei attested
Appearance/result seem, appear, turn out to be NP only / no clear non-NP TFRs
Marginal/unclear happen to be mostly standard FRs; a few ambiguous NP cases

Follow-up corpus checks

  • Seem / appear with PP or AdvP yielded only standard FRs or measure/degree cases, not clear category-transparent TFRs.
  • Turn out to be currently behaves like the seem / appear group, not like the attributional verbs.

The split is not just AP vs. non-AP, and not just copula vs. no copula. Category flexibility clusters with attributional verbs.

From one construction to an inheritance hierarchy

Working hypothesis: the corpus split is easier to state as inheritance than as one uniform TFR construction.

TFR parent
  • exocentric combination of a tran-free-rel clause with a predicative nucleus
  • agreement transparency and NP nuclei available across the broader verb set
  • no commitment yet to a single constructional CI for every verb type
Attributional daughter
  • verbs such as call, consider, describe as, regard as, take to be
  • full category transparency: NP, AP, AdvP, PP
  • CI = attrib-frame with SOURCE and PRED

Seem / appear and perhaps turn out to be belong in a different subtype: same exocentric syntax, but no source-attribution CI and no non-NP nuclei.

Parent construction and attributional daughter

Trans-Free-Rel-Cl construction

S
tran-free-rel
GAP<2 XP[PRD +]>
\(\rightarrow\)
1 NP[FORM what] +
S
VFORMfin
GAP<1 NP, 2 XP[PRD +]>

Attributional TFR daughter

XP
tfr-cxt
CI
attrib-frame
SOURCE3
PRED4
\(\rightarrow\)
S
tran-free-rel
GAP<1 XP>
H 1 XPs

For the attributional subtype, attrib-frame works better than nonspk-commit. The parent construction states the exocentric syntax; the daughter adds source attribution and full category transparency.

Worked example

Breast cancer is (what scientists call) multifactorial.

Phrase-level AVM

AP
tfr-cxt
CI
attrib-frame
SOURCEscientists
PREDmultifactorial

Clause plus nucleus

S
tran-free-rel
GAP<1 AP[PRD +]>
realization: what + scientists call ___ + AP nucleus multifactorial

The nucleus AP determines phrase-level category. SOURCE = scientists; PRED = multifactorial.

Why non-commitment is too strong

Hard-coded CI

TFRs contribute nonspk-commit(s), which predicts speaker distancing across the board.

Counterexample

“This is what I call the ‘transparency effect.’”

That is endorsement, not hedging. The CI cannot hard-code non-commitment.

From source attribution to commitment

The construction contributes an attributional frame. Commitment is computed downstream from the source relation.

Form Result
what I call X endorsement
what they call X hedging
what is rumoured to be X evidential distancing
Subtype consequence

Seem / appear TFRs are related, but they lack an identifiable source and remain NP-only.

CI projection

  1. It is not true that breast cancer is (what scientists call) multifactorial.

    CI projects: “scientists apply this label” persists under negation.

  2. If breast cancer is (what scientists call) multifactorial, then ...

    CI projects from the conditional antecedent.

  3. Is breast cancer (what scientists call) multifactorial?

    CI projects from the question.

The attributional frame is constructional. The specific source is compositional, from the embedded subject.

What does this predict that Grosu (2003) doesn’t?

  1. CI content: feature percolation handles structure, but it does not provide a principled place for the attributional-frame CI.
  2. Verb restriction + CI coupling: in Grosu, verb restriction and CI are separate stipulations. In SBCG, the verb’s argument structure and the CI are coupled by the same construction.
  3. Two kinds of transparency: non-NP nuclei are confined to attributional verbs. Take to be patterns with call, while turn out to be now looks closer to seem / appear.

The CI, its coupling to verb restriction, and the link between attribution and category flexibility are construction-specific claims that Grosu’s account does not derive on its own.

Cross-linguistic and future directions

Cross-linguistic

Korean kes and Japanese koto / no show related patterns.

German may lack productive TFRs because it lacks the right verb + predicative-complement + attribution configuration.

Test program
  • speaker-source vs. other-source experiments
  • evidential gradient across verb types
  • boundary of the attributional class: turn out, happen, prove

Morphosyntactic and semantic transparency typically co-occur, but they are independently motivated.

Polyfunctionality as constructional licensing

Main claim: what’s range is licensed construction by construction, not by a single lexical entry.

TFRs add a new point in the polyfunctionality space, licensed by:

  • structural constraint: GAP + PRD, exocentric
  • lexical constraint: verb inventory, coupled to CI
  • pragmatic constraint: attributional-frame CI

The distributional facts, the verb restrictions, and the CI follow from the construction, not from what itself.

Discussion

TFRs are not one uniform transparency effect.

  • exocentric syntax is general
  • full category transparency clusters with attributional verbs
  • commitment follows from source attribution, not from a hard-coded hedge

Jong-Bok Kim: jongbok@khu.ac.kr
Brett Reynolds: Brett.Reynolds@humber.ca
Slides: https://brettreynolds.ca/wrapp-slides.html#/title-slide