📄 Syntagmatic Productivity of MWEs in Scientific English

Summary

The paper asks whether multiword expressions in scientific English broaden the range of contexts they occur in over time, and whether this broadening outpaces that of comparable single words. The authors use the Royal Society Corpus (RSC) 6.0, which covers 1665 to 1996. They parse it with Stanza and detect five MWE classes, mostly from UD relations plus an academic formula list. Each detected expression is collapsed into one unit. Syntagmatic productivity (SynProd) is then measured as the entropy of a unit's neighbours within three tokens, computed separately for the left side (L3) and the right side (R3). Scores are averaged per decade.

Every MWE class grows more productive over time, and the growth speeds up from roughly the mid-19th century. MWEs also gain more than single-token units of similar parts of speech. Fixed expressions and academic formulas end up above function words. The authors read this as evidence that formulaic language in science becomes increasingly conventionalized. This extends earlier SynProd work, which covered parts of speech only up to 1920 and ignored MWEs.

Key points

• Data: RSC 6.0 contains 47,837 texts and 295,895,749 tokens, taken from two journals of the Royal Society of London: its Proceedings and its Philosophical Transactions.
• Parsing: the corpus was parsed with Stanza's combined English model, trained on the EWT, GUM, GUMReddit, PUD and Pronouns UD corpora.
• MWE types: 3,147,703 types were found in total:
  • compound: 2,523,696
  • flat: 604,057
  • phrasal verb: 16,337
  • fixed: 3,107
  • AFL (Academic Formulas List): 506
• Class trajectories:
  • Compounds have the lowest SynProd until about 1730 and then approach the level of nouns.
  • Phrasal verbs stay below the all-token average until the mid-19th century but overtake it in the last decades of the 20th century.
  • Fixed and AFL expressions exceed function words after the mid-19th century. AFL scores higher than fixed on R3, but the two are similar on L3.
• Acknowledged limitation: sub-corpus sizes per period were not balanced, which can distort entropy. Discipline mix over time was not controlled either.

Interesting ideas

• I1. Detected MWEs are fused into one token using a separator that never appears in the corpus ("||"), so ordinary per-token statistics apply to them unchanged.
• I2. MWE classes are read straight off UD relation labels (compound, flat, fixed) plus particle-verb pairs, so the parser does the MWE detection with no separate extractor.
• I3. Context entropy is measured separately on the left and right sides, revealing asymmetries such as formulas being more varied in what follows them than in what precedes them.
• I4. Each MWE class is judged against a matched single-token baseline (phrasal verbs vs. verbs, compounds vs. nouns, formulas vs. function words) rather than in isolation.
• I5. The authors note that unequal period sizes can inflate entropy and plan equal-sized samples instead.

For Mova

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

• I1 (tokens): the lemmatizer could emit reversible fused units for fixed expressions like "because of" or "in spite of", letting later stages treat them as single function words without losing the original split.
• I2 (tree): gparse output could feed an MWE inventory directly, and disagreement on these labels across witness parsers would flag unstable multiword spans.
• I3 (world): directional neighbour entropy could serve as a cheap, explainable signal of how fixed or open a slot is, useful when deciding whether an expression names an entity or plays a functional role.
• I4 (learn): evaluations of multiword handling could always report a matched single-word control, so improvements are not mistaken for general corpus drift.
• I5 (learn): any entropy or diversity metric over silver data should be computed on size-matched samples as a negative control against corpus-size effects.