📄 Emerging English Transitives over the Last Two Centuries

Summary

The paper asks how English acquires new transitive verbs and tests two hypotheses on that question. The data is the British Hansard, two centuries of UK parliamentary debates, parsed automatically with Stanford CoreNLP. A rule layer on top of the dependency output labels each verb token as transitive (has a dobj), passive, or without an object, and filters out noisy lemmas. The authors compare an early bin (1800–1849) with a late bin (2000–2003); the two have similar word counts. Using χ² tests with Bonferroni correction, they find 196 verbs that are transitive in the late period but absent in the early one.

They then test two hypotheses:

• Rarity and change. Rarer lemmas do change more readily, but only up to a point. The very rarest lemmas change less than moderately rare ones.
• Delayed passive. Logistic growth curves are fitted in Stan to the active and passive frequencies of each new verb. The passive often lags the active, but not always. Some verbs gain both voices together, and no verb clearly gets its passive first.

The work shows that parsed historical corpora can be used to measure syntactic change, not just lexical change.

Key points

• Corpus: British Hansard, 1803–2003, over seven million speeches. The early bin (1800–1849) holds about 51.6M words and the late bin (2000–2003) about 66.3M.
• Significance testing: 2864 transitive verbs had more than 10 occurrences across the two bins. With a Bonferroni-adjusted threshold of 0.01/2864, 1439 of them show a significant change.
• New transitives: A verb counts as new if it appears at least 20 times as a transitive in 2000–2003 and never in 1800–1849. This yields 196 verbs. Topping the frequency list are highlight and monitor, followed by outline, chair and clarify, and then contact, target, underpin, envisage and stress.
• Cleanup filters:
  • Lemmas used as verbs less than 10% of the time are dropped.
  • Tokens with special characters, fewer than three letters, or no vowels are discarded.
  • Hyphenated and unhyphenated spellings are merged.
• Logistic model: Each verb gets a logistic fit over T = 203 years in Stan. Verbs are excluded when:
  • the passive's final rate is below 20% of the active's;
  • either fit has high error;
  • the steepness k is too small to locate an inflection point.

Interesting ideas

• I1. When two coordinated verbs share an object but the parser attaches it only to the first, copy the dobj arc to the second verb as well.
• I2. The authors show why a tagger with 1% error is useless for rare usages: when a lemma is 99% noun, the mistagged noun tokens can equal the true verb tokens. Their fix is to require a minimum verb share per lemma before trusting verb readings.
• I3. Treat npadvmod measure phrases ("eased a fraction") as non-objects, and ignore verbs that a noun governs unless they head a relative clause (e.g. "ticketing machines").
• I4. Compare corpus slices of equal size rather than equal time span, using per-lemma contingency tables, χ² tests and multiple-comparison correction. The authors also caution that a significant shift may reflect changing topics rather than language change.
• I5. Before reading meaning into a curve's parameters, fit the logistic and reject cases where the fit error is high, the growth is too flat, or one series is too sparse.

For Mova

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

• I1 (tree): Add this as a post-parse repair rule so that in "stolen and leaked documents" both verbs get the object, which keeps frames complete.
• I2 (tokens): Store per-lemma POS-share priors and treat minority-POS readings of noun-dominant lemmas like "book" with suspicion before they reach the lexicon.
• I3 (frames): Do not turn such NPs into patient roles. Map measure phrases to an extent or degree slot instead of who-did-what-to-whom.
• I4 (learn): Use this to detect which verb argument frames differ between silver data from different genres. A Bonferroni-style threshold plus a topic-shift negative control would keep false alarms down.
• I5 (learn): Apply the same gating when tracking learning curves or the adoption of repair rules, so that conclusions come only from curves that really have a clear inflection.