Zipf at the frame: English verb-argument constructions vs organic-chemistry frameworks,
What’s claimed and how well supported
- Labels on this page
- SPECULATIVE ×2
- Strongest label used
- SPECULATIVE
- Main caveat (quoted)
- Setting them side by side is a ◆ SPECULATIVE comparison inside Grant's frame, not a correlate and not a joint field.
Counted automatically from the labels printed on this page. Counts are a reading aid, not a grade of the page.
Terms on this page (1), plain-language definitions
- Dopant
- Intentional dilute impurity meant to tune electronic/optical/catalytic behavior without replacing the host as majority phase.
Definitions quoted from the Chem for linguists primer glossary.
Date: 2026-10-06 PT (Navajo lexicalization push, early morning)
Status: Scout. The linguistics (construction-grammar corpus work) and the chemistry (CAS Registry framework analysis) are each same-domain description from the sources opened below. Setting them side by side is a SPECULATIVE comparison inside Grant's frame, not a correlate and not a joint field.
Audience: Grant (linguistics; less chemistry).
Does not: claim that a molecular framework is a phrase, a construction, or a theta frame; supply a stand-in for a theta role; invent Navajo forms, glosses, counts or slopes; claim a Zipf law over molecules as species; revise Bridge A.
Why this note. Grant's brief (2026-10-04/05): Zipf at phrase and at molecule is the shared level, and the search there is for non-traditional arguments (the thing that picks a slot), not for lexical items or atoms. The Zipf note (Zipf) left HIS open question: does power-law behaviour attach to verb-plus-arguments units rather than phrases in general? It also said no opened chemistry source showed a power law over whole molecules. This run opened the one linguistics literature that measures Zipf inside an argument frame and the one chemistry literature that measures a power law over the skeletons of tens of millions of molecules.
What is new (one sentence). Both literatures have a sourced power law at the level of a frame, measured with the same kind of cumulative log–log plot and landing near the same value — the verbs filling one English verb-argument construction (V across n: slope −1, R² = 0.984; V about n: −0.687, R² = 0.991; Römer, O'Donnell & Ellis 2015) and the frameworks of 24 million CAS organic compounds (slope −1.07, exponent 2.07; Lipkus et al. 2008) — but they rank opposite things: English ranks the fillers of a fixed, role-ordered frame, while chemistry ranks the frames themselves after deleting every filler and even the attachment points, so neither literature has yet measured the other's half.
1. Sources opened this run
| Source | How read | Used for |
|---|---|---|
| Römer, U., O'Donnell, M. B. & Ellis, N. C. 2015, "Using COBUILD grammar patterns for a large-scale analysis of verb-argument constructions," in Hunston festschrift (Benjamins), ch. 2 | Author-hosted PDF; text layer; page image p. 55 (Fig. 10) checked because the text layer scrambles which slope goes with which panel | Method (CCDF with log binning, Adamic & Huberman 2002); Fig. 10a V across n R² = 0.984, γ = −1; Fig. 10b V about n R² = 0.991, γ = −0.687; lead verb prototypical and generic (COME, TALK); selectivity (χ², 1−τ, entropy); faithfulness and MI (GIVE vs LEAVE in the ditransitive) |
| Ellis, N. C., O'Donnell, M. B. & Römer, U. 2014, "The processing of verb-argument constructions is sensitive to form, function, frequency, contingency and prototypicality," Cognitive Linguistics 25(1): 55–98 | U. Michigan offprint PDF, text layer (pp. 55–61 read closely) | 20 VAC frames (Table 1); four-step method; summary of Ellis & O'Donnell 2011/2012 findings (Zipfian verb island, lead verb prototypical, selective, semantically coherent); Hare & Goldberg priming result as reported here |
| O'Donnell, M. B. & Ellis, N. C. 2010, "Towards an inventory of English verb argument constructions," NAACL HLT 2010 workshop (ACL Anthology W10-0802) | PDF text layer | Zipf as pf ~ f^−b, b ≈ 1; Goldberg, Casenhiser & Sethuraman 2004 (as reported); V across n and ditransitive V Obj Obj both fit straight lines (R² = 0.993); authors' own caution that binning may over-smooth; plan for yoked random controls against "uninteresting artifact" objection |
| Lipkus, A. H., Yuan, Q., Lucas, K. A., Funk, S. A., Bartelt, W. F., Schenck, R. J. & Trippe, A. J. 2008, "Structural diversity of organic chemistry. A scaffold analysis of the CAS Registry," J. Org. Chem. 73: 4443–4451 | Publisher PDF (downloaded as guest), text layer, pp. 4443–4451 | Framework definition and pruning algorithm; three levels (graph; graph/node; graph/node/bond); power law p(x) = Cx^−α, cumulative slope −1.07, α = 2.07; 143 graph frameworks describe half the compounds; hetero frameworks per graph framework top-heavy but not a power law (cutoff); Simon/Yule word-stream model applied to frameworks; synthetic-cost explanation |
| Lipkus, A. H. et al. 2019, "Recent changes in the scaffold diversity of organic chemistry as seen in the CAS Registry," J. Org. Chem. 84 (doi 10.1021/acs.joc.9b02111) | Publisher full-text HTML | Power law holds at two time points 10 years apart; rank order of frameworks fluctuates; cumulative-advantage model alone cannot explain the fluctuations; long tail (singletons) large and growing |
Files under lit/_pdfs/zipf-frames/. Reported only through the papers above, not opened: Goldberg, Casenhiser & Sethuraman 2004; Ellis & Ferreira-Junior 2009; Hare & Goldberg 1999; Ellis & O'Donnell 2011/2012; Solé et al. 2005; Simon 1955; Bemis & Murcko 1996/1999; Karakoc et al. 2006 (small scaffold sets).
2. Language side: Zipf inside a verb-argument construction
What a VAC is here. A verb-argument construction is a frame like V across n ("she ___ across the field"), V about n, or the ditransitive V Obj Obj, defined first by form (from COBUILD Grammar Patterns) and searched in a dependency-parsed 100-million-word British National Corpus. The verb slot is left open.
What is ranked. For one frame at a time, the verb types that fill the verb slot, by token count (e.g. V across n: come 628, walk 243, run 202, cut 198 … veer 3, rip 3 … bump, hare, whir 1; Römer et al. 2015 p. 55).
The measured law. Plotted as log cumulative probability against log frequency, with logarithmic bins: - V across n: R² = 0.984, slope γ = −1 (Fig. 10a, read from the page image). - V about n: R² = 0.991, slope γ = −0.687 (Fig. 10b, read from the page image). - V across n and V Obj Obj in the 2010 pilot: both fit straight lines, R² = 0.993.
The 2015 text says both 10a and 10b have "slope (γ) around 1"; the figure shows −0.687 for V about n. Recorded as printed, not reconciled here. O'Donnell & Ellis (2010) say they will check other fitting methods "to ensure we have not smoothed the data too much through binning."
Three findings that go past "it's Zipfian" (sourced). 1. Selective, not inherited. The authors themselves call a bare Zipf curve "potentially trivial" because Zipf is everywhere in language. What makes it interesting is that the verb ranking inside a frame cannot be predicted from the verbs' overall frequency in English (χ², 1−Kendall τ). Ranked by association instead of raw count, the top of V across n is scud, flit, slant, scuttle, skid, not come, walk, run. 2. The head of the curve is the frame's meaning. The most frequent verb is prototypical of the construction's meaning and generic in its action semantics (COME for V across n, TALK for V about n; put in verb-object-locative, give in the ditransitive). Learners pick that verb up first (Ellis & Ferreira-Junior 2009, as reported). 3. Role order, not constituent structure, is what speakers carry over. Ellis et al. 2014 (p. 57–58) report Hare & Goldberg's priming result: provide-with sentences have the caused-motion constituent structure (NP [V NP PP]) but the ditransitive order of rough semantic roles [Agent Recipient Theme], and they prime ditransitive descriptions as much as ditransitives do, with no priming of caused-motion. Ellis et al.'s conclusion: "when order of semantic roles is contrasted with constituent structure, only the order of semantic roles shows priming." Opened second-hand only.
Point 3 is the sourced piece that matters most for Grant's brief: the frame whose filler distribution is Zipfian is a role-order object, i.e. built out of non-traditional arguments, not out of particular words. That is description of English; it does not by itself tell us anything about chemistry.
3. Chemistry side: a power law over molecular frameworks
What a framework is. Lipkus et al. take every non-polymer, single-component, cyclic organic compound in the CAS Registry with an explicit atom-level structure (more than 24 million). They repeatedly prune terminal atoms until only the ring systems and the chains linking them are left. What remains is the framework. Side chains are gone, and so is any record of where they were attached or by what kind of bond; stereochemistry, charge and 3-D shape are also dropped (p. 4444).
Three levels. The same framework is stored at three grains: the graph level (connectivity only, ignoring which element and which bond type), the graph/node level (adds elements), and the graph/node/bond level (adds bond types) (p. 4445). The graph level is the one without atom identity.
What is ranked and the measured law. For each "hetero" framework (elements and bonds kept), the number of compounds built on it. The cumulative plot (number of frameworks occurring that often or more, against frequency) is linear "much further" than the raw plot, and conforms "almost exactly to a power law." Regression through the origin gives slope −1.07, so α = 2.07 for p(x) = Cx^−α; they note only three of twelve listed power laws in the physical and social sciences have a lower exponent, so the distribution is "unusually top-heavy" (pp. 4447–4448).
At the graph level (no atom identity). Even more top-heavy: 143 graph frameworks (0.017% of them) describe just over half the compounds; 393,144 graph frameworks occur in a single compound (p. 4449).
Fillers of one shape are capped. The number of hetero frameworks (element patterns) per graph framework is top-heavy (1.0% of shapes give 38.5% of hetero frameworks) but "does not appear to follow a power law": its tail has a cutoff, which they attribute to stability, synthetic accessibility and framework size (p. 4450).
Their mechanism, and the word analogy they draw themselves. Lipkus et al. explicitly carry Simon's word-frequency model over: read the Registry as a stream of compounds in order of registration, replace each compound by its framework, and you get a stream of new and old frameworks "analogous to the previous example of words in a text"; the more a framework has been used, the more likely it is to be used again, plausibly because known derivatives lower synthetic cost (p. 4448). So the framework plays the role of a word type in Simon's model; the paper does not say molecule ↔ phrase. The 2019 follow-up finds the power law held across ten years while the rank order of frameworks shifted, and says a pure cumulative-advantage model cannot explain those fluctuations.
This partly corrects the Zipf note's §4 line "No opened full text maps linguistic Zipf onto molecules": Lipkus et al. 2008 do map the Simon/Zipf word model onto molecular frameworks, at the level of a type/token growth process, not as a phrase ↔ molecule claim.
4. Side by side (analysis, not a source claim)
Both are cumulative log–log plots of "how many types occur at least this often." On Moreno-Sánchez et al.'s relation (Zipf note §1: frequency-spectrum exponent β = 1 + 1/α, rank exponent α), a cumulative slope of −1 corresponds to a rank exponent of 1. So, converting only the printed slopes:
| Object | What is counted | Cumulative slope | Rank exponent implied |
|---|---|---|---|
| English V across n | verbs filling one frame | −1 | ≈ 1 |
| English V about n | verbs filling one frame | −0.687 | ≈ 1.46 |
| CAS hetero frameworks | frames, by molecules built on them | −1.07 | ≈ 0.93 |
The conversion is arithmetic on printed numbers; fitting methods differ (log binning vs regression through the origin on the unbinned cumulative), so the closeness of −1 and −1.07 is suggestive, not a match.
The asymmetry is the main result. - English: the frame is held fixed and kept with its slots in order; the fillers (verbs) are what follow the power law, and that law is selective to the frame. - Chemistry: the fillers (side chains) are thrown away together with their attachment points; the frames are what follow the power law. When the closest available filler count is taken (element patterns per shape), it is capped, not a power law. But that is a count of distinct patterns per shape, not a token distribution of fillers inside one frame, so it is not the measurement the VAC papers make. - So the like-for-like tests are both missing from the opened sources: (a) a frame-level ranking in language (how many tokens each VAC gets across the whole construction inventory), and (b) a filler-within-frame ranking in chemistry (which substituents fill a given position on a given framework, by count).
What this does for Grant's non-traditional-argument search. The VAC literature supplies a sourced Zipf object defined by role order rather than lexical identity (§2 point 3). The CAS graph-level frameworks supply a sourced power-law object defined without atom identity. Those are the two places in the opened literature where a power law sits on something other than the words or the atoms. But the chemistry object is a shape, not a frame with ordered slots, because the attachment points are deleted. A chemistry object with slots (positions on a framework where substituents go) is what would have to be ranked to make the comparison like for like.
5. Navajo
Nothing opened this run has Navajo counts. Chee 2025 (as recorded in Phrasal prescription) describes the Navajo phrasal verb as a postposition plus a minimally inflected verb, which is formally the same kind of object as an English V across n frame (a relational word plus an open verb slot). Whether the verbs or verb themes that fill a given Navajo postposition frame are Zipfian, selective and headed by a generic prototype is an empirical question that needs a Navajo corpus or Grant's texts. Not extrapolated. The causative-split question (Japanese-style lexical vs derivational) is untouched by these sources.
6. Separation
SOURCED (do not inflate). Zipfian, frame-selective verb distributions in English VACs with the printed slopes; lead verb prototypical and generic; role-order priming as reported; power-law framework distribution in the CAS Registry with α = 2.07; 143 graph frameworks for half the compounds; per-shape element patterns cut off; Simon's model carried over by Lipkus et al.; rank-order fluctuation over ten years.
HIS SPECULATIVE (not findings). Phrase ↔ molecule, word ↔ atom; "phrasisized" units as phrase-level requirements for a molecule; a Navajo phrasal verb as a prescription with dopantIntentional dilute impurity meant to tune electronic/optical/catalytic behavior without replacing the host as majority phase.-like places.
Not claimed. That a framework is a phrase, a VAC, or a theta frame; that side chains are arguments; that the slope match means a shared mechanism (the papers offer different ones: learnability for VACs, synthetic cost and cumulative advantage for frameworks); any Navajo frequency fact.
7. Open questions
- Chemistry, filler-within-frame: open Ertl's substituent analyses (e.g. "Cheminformatics analysis of organic substituents," 2003) or an R-group frequency study for a fixed scaffold, and check whether substituents at a given position on a given framework are power-law distributed and frame-selective, as verbs in a VAC are. Not opened.
- Language, frame-level: is the token count per VAC across the whole COBUILD-derived inventory itself Zipfian? Solé et al. 2005 is cited for "grammatical constructs" following Zipf; not opened.
- Read Goldberg, Casenhiser & Sethuraman 2004 and Hare & Goldberg 1999 directly (only reported here).
- Benz, Swamidass & Baldi 2008 (ring systems, rigid segments; abstract only in the Zipf note) against Lipkus: same object or different cut?
- Navajo: once Grant's texts or a corpus are in hand, count verb themes per postposition frame (Chee 2025's phrasal verbs) and run the VAC tests (Zipf fit, selectivity, generic prototype).
- Multiple-wh / ranking-edge tests remain with his texts; nothing here bears on them.
Follow-up pointer, 2026-10-06 midday
Open question 1 is taken up in Zipf fillers in chemistry (site page: Zipf fillers in chemistry). Substituent counts are long-tailed (Ertl 2003; Ertl 2021 preprint), but no opened source prints a slope or gives a per-position distribution on one framework, so the like-for-like test is still missing. The nearest slot-level object is benzene's substitution patterns with filler identity ignored (Nilova & Stuart), which those authors tie to the directing rule. The gap is narrowed, not closed.
Corpus note: zipf-frame-filler-vac-scaffold-v0.md. Rank counts stay in the field that measured them. Grant’s phrase↔molecule line, phrasal-verb question, and quark/anyon analogy are his. Not Bridge A. Not a joint field.