📄 Drivers of English Syntactic Change in the Canadian Parliament

Summary

The paper asks why the English of the Canadian House of Commons became more conversational over the twentieth century. It considers three possible causes: younger members replacing older ones (cohort replacement), women entering Parliament, and individual speakers changing their own style. The authors parse the speaker-labelled Canadian Hansard (1901–2017) with Stanford CoreNLP. They then count five syntactic markers of colloquialization per verb phrase: be-passives, progressives, modals, semi-modals and pied-piped relative clauses. These counts are turned into Z-scores and combined into a single colloquialization score. Multilevel regression, including a latent-variable model, links the score to year, age, seniority, gender, party and being in government. The results rule out the two popular explanations. Generational turnover accounts for only about a tenth of the trend, and women speak somewhat more formally than men. Holding power makes speakers markedly more formal. The trend is driven mostly by individuals growing more colloquial over their careers. This matters because it shows that tracking individual speakers over a long period can separate causes that aggregate corpus counts mix together.

Key points

• Data: the Canadian Hansard has over 650 million words covering 1901–2017, with 1906–1907 missing. The authors used version 1.0.2 from LiPaD. Each speech is linked to a member ID with biographical details.
• Feature detection: features are found with CoreNLP dependencies, POS tags and lemmas.
  • Passives are identified by auxpass/nsubjpass.
  • Progressives are a form of "be" followed by a VBG.
  • Pied-piping is an IN tag directly before who/whom/whose/which.
  • Modals are MD-tagged words checked against a fixed lemma list.
  • Semi-modals (have got to, be going to, want/have/need to) are matched with small lemma-and-word patterns.
• Correlated features: among the top 100 speakers (each with at least 350,000 words), the five features correlate strongly with the expected signs. For example, passives correlate +0.795 with pied-piping and −0.793 with semi-modals.
• Power and party: over 1961–2010, the governing party spoke roughly as formally as speakers did about 6 years earlier, and this reverses when a party loses power. The New Democrats, who never governed, averaged a score of 0.48 and were more colloquial than both main parties.
• Gender and cohort: without controls, women appear about 4.5 years "behind" men in 1961, shrinking to 1 year by 2010. Once power is controlled for, the gap is a stable ~1.6 years. Cohort effects explain about 0.1 of the observed ~1 per year increase, so roughly 90% of the change comes from other factors, mainly individual change.

Interesting ideas

• I1. Calibrate on well-sampled sources: feature rates are normalized using baselines computed only from heavily sampled speakers, so noisy low-volume sources do not inflate the variance estimate.
• I2. Detect constructions with transparent tree rules: colloquialization markers are found with small, readable rules over dependency labels, POS tags and lemmas rather than a learned classifier, so every count can be traced back to a match.
• I3. Report effects in an intuitive unit: regression coefficients are translated into "years of change," which makes effect sizes such as party power or gender easy to compare and explain.
• I4. Re-test an effect after adding a confounder: the apparent narrowing of the women–men gap mostly vanishes once party-in-power is included, a useful reminder to re-test any effect after adding a plausible confounding variable.
• I5. Treat co-varying cues as one latent dimension: several surface features that move together are treated as signals of one hidden variable, formality.

For Mova

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

• I1 (learn): style or construction-frequency baselines used in evaluation could be calibrated only on long, well-attested documents before short texts are scored against them.
• I2 (tree): similar explainable detectors over UD trees (using aux:pass/nsubj:pass after dialect conversion) could flag passives, progressives and pied-piped relatives for later stages.
• I3 (learn): improvements from repair rules or teacher tiers could be reported in a comparable intuitive unit, such as the equivalent amount of extra training data, rather than raw deltas.
• I4 (learn): evaluations could routinely add confounding controls (genre, sentence length) to check that a claimed parser or reader gain does not disappear.
• I5 (frames): pragmatic analysis could estimate a register or formality variable from modals versus semi-modals ("shall" versus "have to"), passives and progressives, and use it when interpreting speech acts and politeness.