LIT-410: The English phrase-as-lemma construction: When a phrase masquerades as a word, people play along

Goldberg and Shirtz (2025) — Language 101(2):291-320, https://doi.org/10.1353/lan.2025.a962899

Key takeaways

• The phenomenon is a unit with the internal syntax of a phrase in a slot reserved for a word, and it goes up to whole sentences: a trickle-down policy, a must-do task, an 'I'm not a witch' moment, a 'take music for granted' attitude. All attested in COCA.

• It resists every familiar category, and the argument is distributional rather than theoretical. Against a compound analysis: very and more modify PALs (a very 'I'm white and right' way, a much more 'I'm compassionate too' campaign), which modify adjectives rather than compound nouns; and PAL + Noun cannot scope over a distinct head noun, which NN and AN compounds can. For one: PALs take compound stress (an 'I'm not a WITCH' moment, not a nonwitchy MOMent) and bind tightly enough that adjectives resist intervening. The authors’ conclusion is that the construction needs its own representation — external syntax of a word, internal syntax of a phrase — rather than assimilation to Noun or Adjective.

• Speakers act on the word-slot, and that is the measured part. Four preregistered surveys on Prolific, against 100 sentence pairs normed for semantic similarity (target pairs at M = 90.4, statistically indistinguishable from near-synonymous fillers at 93.0). 700 recruited, 685 after two catch trials:

  • shared common knowledge — PAL chosen 77.3% (β = 1.69, p < 0.0001); a separate ‘shared background’ wording, 74.3%
  • wittiness — 82.2% (β = 2.58, p < 0.001)
  • sarcasm — 84.5% (β = 2.71, p < 0.001)

• No frequency effect, and Study 4 is the reason that matters. Estimated log COCA frequency of each PAL phrase predicted nothing. A fourth survey (70 new participants) restricted to high-frequency phrases — anything goes, all you can eat, do it yourself, wait and see, trickle-down — reproduced every effect at roughly the same size (common knowledge 72.9%, wittiness 79.8%, sarcasm 84.5%). So the effect is not memorization of particular strings. It is carried by the construction, which is what makes it a claim about structure rather than about a list.

• The mechanism proposed is lemma-like meaning. A word-slot implies a lemma; a lemma presupposes familiarity with the semantic frame it names (Fillmore); so putting a phrase there asserts that the situation it describes is one the listener already recognizes — the linguistic analogue of observational humor. The form’s peculiarity and its function are claimed to be the same fact, which is why the paper predicts comparable constructions in unrelated languages (Hebrew and Brazilian Portuguese are observed).

• Stated limitations. English only, with cross-linguistic comparison left to future work and flagged as hard because reference grammars omit infrequent constructions. The analysis is restricted to phrases containing a verb; other phrase types, non-English quotes and noises (kids are 'doo-doo-doo-doo'-ing the day away) are expected to assimilate but were not tested.

Standing in the anthology

It sources THEORY-022 and no practice, and both halves of that are deliberate.

The theory is the claim the surveys measure: a phrase in a word slot is read as a lemma, and the reading comes from the construction rather than the phrase’s frequency. That is a finding about what is true, which is what the THEORY scheme holds — and it is a finding about the signal, not about the people who were asked. A claim discovered in psycholinguistics still takes its kind, the same way ADR-026 has a preconditioning scheme found in diffusion take its kind rather than its domain.

The practice is absent because no recommendation follows yet. Knowing that the data contains construction-delimited units of arbitrary length says nothing about whether a vocabulary should try to hold them; that is a cost question, and nobody in this record has measured it. Filing a practice here would be the anthology’s inference wearing this paper’s citation, which ADR-017 is the rule against.

What the record is missing to connect it. The bridge from this finding to model internals is the detokenization and inner-lexicon line — ARXIV-2410.05864, ARXIV-2406.20086 — which reports models fusing multi-token words into single representations in early layers. The record holds none of it, and until it does the connection between this measurement and anything a network does is a conjecture rather than a claim.

The tag is the closest available and still does not fit cleanly. representation-and-encoding is about encoding for a network, and this is about the structure being encoded. That is a finding about the vocabulary rather than about the paper — DP-009 — and saying so rather than absorbing it quietly is what DP-008’s corollary requires. ADR-045 puts the question.