Cambridge 2026 · UK headlines · vocabulary

Novel words in UK test

How many UK-test words never appear in UK training — and how many never appear in AG training? Reported as both types (unique words) and tokens (occurrences). Out-of-vocabulary tokens are one concrete form of domain / split shift.

vs UK train: 602 types · 739 tokens (17.7%)
vs AG train: 200 types · 255 tokens (6.1%)
UK test: 1,788 types · 4,165 tokens

Headline counts

A token is novel if that surface form never appears in the comparison training corpus. Types = unique novel words; tokens = how often those novel words occur in UK test. Tokenization: lowercase whitespace split (same as the Lengths page).

Comparison Novel types % of types Novel tokens % of tokens
UK test − UK train 602 33.7% 739 17.7%
UK test − AG train 200 11.2% 255 6.1%
UK test − both trains 160 8.9% 191 4.6%
Why AG covers more UK test words than UK train does: AG train is much larger (28,500 headlines, 38,838 word types) than UK train (1,216 headlines, 4,656 types), so many everyday English tokens in BBC test already appear in AG — even when UK-specific names and politics terms do not. Token rates are lower than type rates because novel words are often rare.

By label

Same novelty definition, sliced by UK test label. Each row uses only that label’s UK test headlines, still compared against the full UK / AG training vocabulary. Both types and tokens are shown.

Label n Tokens Types vs UK Tokens vs UK % tokens Types vs AG Tokens vs AG % tokens
business 76 1,215 178 218 17.9% 58 68 5.6%
sport 75 1,195 185 224 18.7% 74 97 8.1%
tech 50 796 137 151 19.0% 22 23 2.9%
politics 60 959 120 146 15.2% 47 67 7.0%
Sport is the most AG-novel; tech the least — in tokens too. Sport has 97 novel-vs-AG tokens (8.1% of its tokens); tech only 23 (2.9%). Type rates look higher because many novel words appear only once.

Examples by label (novel vs AG)

Frequent alphabetic UK-test words for each label that never appear in AG train.

business

  • deccan
  • eurodisney
  • rupee
  • sell-offs
  • botin
  • daiwa

sport

  • mirza
  • sania
  • fly-half
  • gallas
  • hodgson
  • extra-time

tech

  • posters
  • adaptor
  • anti-tremor
  • reboot
  • funkier
  • kyrgyz

politics

  • tory
  • ukip
  • eurosceptic
  • auschwitz
  • jowell
  • rebate

Vocabulary sizes

Unique lowercase whitespace tokens per split.

Split Headlines Vocab size
UK train 1,216 4,656
UK test 261 1,788
AG train 28,500 38,838

Shared with UK test: UK train covers 1,186 types; AG train covers 1,588.

Overall novel types vs tokens

UK-test words missing from each training set — unique types and total token occurrences.

Example novel words

Frequent alphabetic examples from UK test (punctuation-heavy tokens omitted). Full lists live in artifacts/novel_words_summary.json.

Novel vs UK train

  • ford
  • allegedly
  • embracing
  • expects
  • half-life
  • henry
  • posters
  • absa
  • ajax
  • arnesen

Novel vs AG train

  • tory
  • ukip
  • mirza
  • sania
  • auschwitz
  • eurosceptic
  • fly-half
  • gallas
  • hodgson
  • barkley