📄 English verb regularization in books and tweets

Summary

The paper measures how far English speakers have moved from irregular past-tense forms (e.g., 'burnt') toward the regular -ed forms ('burned'). It also asks whether edited published text tells a different story from everyday writing. The authors used the verb list of Michel et al. and two sources. The first is Google Books Ngrams (English Fiction, American English and British English 2012 corpora, averaged over 2003–2008). The second is a roughly 10% random sample of Twitter from 2008–2017. For each verb they computed a regularization fraction: the share of past-tense tokens, preterite and participle together, that use the regular form. They compared books with tweets, US with UK, and US counties, and checked county values against census variables. Twitter as a whole regularizes more than fiction books. US-geotagged tweets regularize more than American books, but UK tweets regularize less than British books. Regularization varies by county, yet once population is accounted for, no strong links to education or income appear. The work shows that unedited social media can reveal language use that editorial norms in books conceal.

Key points

• The Twitter "decahose" contains more than 106 billion messages from about 750 million accounts (Sept 2008 – Oct 2017).
• Subsets used:
  • GPS-geotagged US tweets: about 0.27% of the data.
  • GPS-geotagged UK tweets: about 0.05%.
  • Tweets whose free-text 'city, state' profile location matched a US county: 2.26%.
• A verb counts as "regular" when its regularization fraction exceeds 0.5. The AE vs BE comparison is restricted to verbs whose irregular past ends in -t.
• County analysis drew on 513 socio-demographic and economic variables, sourced from the American Community Survey (5-year estimates, 2015).
• Profile locations were matched to counties with the uszipcode package, after raising its fuzzy-match confidence from a hard-coded 70 to 91. A manual check of 1000 matches (100 per year) found only one ambiguous case, 'Long Island, NY'.

Interesting ideas

• I1. One simple ratio (regular tokens over regular plus irregular, pooling preterite and participle) turns a messy morphological variation into a comparable number per verb, per region or per corpus.
• I2. The authors contrast edited text (books, filtered by gatekeepers) with unedited text (tweets) to expose where formal norms hide actual usage.
• I3. Averaged regularization is plotted against each county's total token count and split into frequency bins, and only counties with at least 40 tokens are kept, which shows how small samples distort aggregate values.
• I4. Fuzzy geographic matching was made trustworthy by tuning its threshold and then hand-auditing a year-stratified sample, which also surfaced the one systematic ambiguity.
• I5. British and American English differ in how much they keep the irregular -t past (learnt/learned, spelt/spelled), so verb forms carry a dialect signal.

For Mova

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

• I1 (tokens): the en lemmatizer could store observed variant ratios per verb and map nonstandard forms like 'seeked' or 'sweared' to the right lemma while flagging them as regularized, rather than treating them as unknown words.
• I2 (learn): evaluation sets should be split into edited and informal text, so that accuracy gains on clean prose are not mistaken for robustness on everyday writing.
• I3 (learn): per-subset metrics such as per-genre or per-construction scores should be reported alongside sample size, with a minimum-count floor, so that noisy small slices are not read as real effects.
• I4 (world): linking text mentions to Wikidata places or entities could use a tuned confidence cutoff plus a stratified manual audit, recording known ambiguous names as explicit exceptions.
• I5 (back): generation could choose past-tense forms consistent with the dialect of the source or target text, using the paper's irregular-verb table (Table A1) as a starting inventory.