📄 The Fate of Complex Languages: Classical Arabic and Old Norse in the Age of Globalisation

Summary

Kusters asks whether the social history of a speech community shapes how complex its verbal inflection stays over time. He treats globalisation broadly, so it also covers older imperial and cultural expansions. He predicts that inflection becomes more transparent and more economic when a community has many second-language learners, frequent dialect contact and migration, and low prestige for its variety. "Transparent" means each meaning maps to exactly one form. "Economic" means few obligatory categories. He tests this on two families, each with one ancestor and three descendants whose social histories differ. Classical Arabic is compared with Najdi, Moroccan and Nubi Arabic, and Old Norse with Icelandic, Faroese and Bokmål Norwegian, using verb paradigms drawn from descriptive grammars. The prediction holds overall. Isolated, prestigious communities such as Najdi and Icelandic keep their complexity, while contact-heavy ones simplify, most strongly in the creole Nubi. Sound change makes forms easier to merge, but which distinctions actually disappear depends on morphology and meaning. The paper matters because it links language structure to social history and suggests that complex inflection survives globalisation mainly when it is deliberately protected.

Key points

• Arabic varieties and their profiles:
  • Najdi (Shammar tribe, northern Saudi Arabia) is isolated and prestigious and remains the most complex.
  • Moroccan Arabic has a long history of Berber second-language speakers and low prestige.
  • Nubi (Uganda) arose in military camps where native Arabic speakers fell from about 20% to 4% of the population.
• Arab migration to North-West Africa came in two waves: about 150,000 people between the 7th and 11th centuries, then about 1 million between the 11th and 14th centuries.
• Changes in Najdi: dual and mood are lost, u and i largely merge, and there are fewer verb classes. Phonological rules hide the underlying vowel patterns that signal voice and aspect. The further a dialect is from the Shammar, the more these vowel distinctions collapse.
• Changes in Moroccan: the stem vowels follow syllable structure only, so aspect lives entirely in the affixes. The prefix tends to mark person and the suffix number, giving a more one-to-one split.
• Order of loss in Scandinavian: person and mood go before number and tense. When Norwegian's past-tense dental suffix weakened between vowels (13th–14th centuries), a new dental marker appeared in exactly the endangered forms.

Interesting ideas

• I1. The paper measures transparency by counting departures from a one-to-one meaning–form mapping (fusion, fission, allomorphy, homonymy), which gives a reusable, interpretable complexity metric for any paradigm.
• I2. The author keeps the underlying form separate from the surface form produced by sound rules, because rules can hide the meaning-bearing pattern.
• I3. The study design holds the ancestor constant and varies the social history across three descendants, a controlled comparison that isolates one causal factor.
• I4. Simplification follows a category hierarchy: person and mood erode before number and tense, and reductions start in marked classes.
• I5. When a distinction is about to vanish, a new marker appears precisely in the threatened forms, so the category is preserved even though its exponent is replaced.

For Mova

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

• I1 (tokens): Score the crate en morphology tables with these four counts to find where English forms are ambiguous (e.g. -s, -ed homonymy) and need context-based disambiguation.
• I2 (tokens): Store the lemma, the applied orthographic or morphological rule and the surface form as an explicit trace, so every lemma decision stays explainable.
• I3 (learn): Build evaluation sets as minimal triples (same base text, one varied factor) to act as negative controls showing which input property actually drives a parser or reader change.
• I4 (learn): Use the same priority when tagger features conflict or data is scarce, protecting tense and number agreement first and treating person/mood errors as lower-cost.
• I5 (learn): Self-induced repair rules could target the exact contexts where a needed distinction collapses (e.g. ambiguous -ed past vs participle), recovering it from another cue instead of relearning globally.