Skip to content
← All projects

2026 · Computational modelling for a master’s thesis on Basque revitalisation

Who speaks Basque first

A simulated Basque Country where every person ages, moves, goes to school and talks to whoever is nearby. Fitted to twenty years of survey data, then used to ask which language policies would actually move the number of speakers.

The model

About a thousand agents, built on Civico (2019) and adapted to the Basque Autonomous Community.

The engine

Schooling in Basque drives almost all of the growth. Everyday conversations drain it.

The lever

Who opens a conversation in Basque matters as much as tripling adult classes.

A sociolinguistic question, a physicist’s tool

Can a simulation say anything about the future of a minority language? This project was the computational part of a master’s thesis on Basque revitalisation. The tool is an agent-based model, standard in sociophysics: you write down how individuals behave, let thousands of them interact, and see what the population does.

I designed and implemented the model, adapted it to the Basque Autonomous Community, calibrated it against official statistics and ran the simulations and policy experiments. The thesis, written in Basque, sets them in their sociolinguistic context; this page tells the computational story.

Everything starts with a conversation

The model starts from Marco Civico’s 2019 study of minority languages, which he validated on Romansh in Switzerland. Everyone speaks the majority language. Some people also speak the minority one, with a fluency between 0 and 100. Each year every person moves a little and has one conversation with whoever is closest. Four things can happen:

The key ingredient is politeness. Speaking Basque to someone who does not know it can feel rude, so many speakers open in Spanish just in case. Civico calls them reserved. If two reserved speakers meet, the conversation happens in Spanish even though both could have spoken Basque. In the calibrated model, 65% of speakers are reserved.

0%25%50%75%100%years0
0% speak Basque
  • No Basque
  • Fluency 1–33
  • 34–66
  • 67–100

Running…

Only the conversation rule, with nobody born, dying, schooled or arriving. Each dot is a person; each year everyone talks once with whoever is closest. Speakers drift towards three Basque-speaking communities. Move the slider while it runs: the break-even point sits near one speaker in two opening in Basque.

Run on its own, the conversation rule behaves like a system with a tipping point. Below roughly half of speakers opening in Basque, more conversations erode fluency than build it, and the language slowly disappears. Above it, Basque spreads. At the calibrated 35%, conversations alone are a net loss. Something else has to keep the language growing.

From Romansh to the Basque Country

Civico’s community is closed and static. Basque is not: people move to Euskadi, children learn the language at school as much as at home, and adults go to classes. I added each of these and tied every parameter I could to official data, mostly the Sociolinguistic Survey, Eustat and HABE.

Migration
Net immigration of 0.6% of the population a year, with the age profile of people arriving in Euskadi (Eustat). Newcomers start with no Basque.
Family transmission
From the Sociolinguistic Survey, 2001–2021: a child learns Basque at home 92% of the time if both parents speak it, 49% if only the mother does and 42% if only the father does.
School
Each child is placed in model A, B or D with the 2006 shares (22.6%, 23.0%, 54.4%). B and D add 6 fluency points a year from age 6 to 15; A adds 1.
Adult classes
Each year 0.3% of the population starts Basque at an euskaltegi, the share enrolling at A1 according to HABE.
Communities
Speakers belong to one of three Basque-speaking areas and drift towards it, so they meet each other more often than chance.
Fluency bands
A 0–100 scale cut into the survey’s categories: no Basque, easier in Spanish, equally easy, easier in Basque. Gains and losses depend on the band.

The demography was calibrated first: deaths, births and migration together have to reproduce Eustat’s population curve, which grew 4.4% between 2006 and 2024. Only then did I turn to the language.

Does it reproduce the last twenty years?

The simulation starts in 2006 with the real age pyramid and the real fluency mix: 30% of people aged 16 and over spoke Basque. The survey has measured that share every five years since, and it has risen to 36%. That is the curve the model has to follow.

Not all the conversation parameters can be measured, so some have to be fitted. To keep the fit honest I fixed the structure of the fluency rules and fitted only three scale factors: how much a Basque conversation adds, how much a Spanish one costs, and how often people without Basque pick some up. I fitted them on the 2011 and 2016 surveys and kept 2021 back as a genuine prediction.

Calibration
25%30%35%40%45%20062011201620212024fittedpredicted
  • Model, mean of 200 runs
  • 80% of runs
  • Sociolinguistic Survey
Share of people aged 16 and over who speak Basque in the Basque Autonomous Community. Three scale factors were fitted to the 2011 and 2016 surveys; 2021 was left out and is a prediction. The bars compare the 2021 fluency mix, survey (blue) and model (amber).

The model runs about a point high throughout and predicts 38% for 2021 against the 36% measured. The survey points stay inside the spread of the runs. The fluency mix is right at the extremes and less right in the middle. The model has too many people who find both languages equally easy.

More important, two quite different calibrations fit the data equally well. One makes conversations in Basque more rewarding (A, talkative). The other makes conversations in Spanish less costly (B, forgiving). Four survey points cannot tell them apart, so every result below is shown for both. Where they agree, the conclusion does not hinge on parameters nobody has measured.

What actually drives the curve

A model that fits is only useful if you know why it fits. The cleanest way to find out is to switch mechanisms off one at a time and re-run 2006–2050.

Calibration
0%20%40%60%2006202020352050
  • Calibrated model
  • No Basque at school
  • Nobody opens in Basque
  • No conversations
  • No transmission at home
Share of people aged 16+ who speak Basque, mean of 200 runs per line. Each line switches off one mechanism and keeps the rest. Hover to read the values. The dashed line sits almost exactly on the calibrated model.

School is the engine. Without Basque at school the share of speakers falls by 20 points by 2024 and almost vanishes by 2050. Nearly all the growth since 2006 is cohorts educated in models B and D replacing older ones.

Conversations are a sink. Remove them altogether and Basque grows faster, by about 5 points by 2024. What keeps them from being a disaster is the minority who open in Basque: if nobody did, the share would collapse by more than 30 points.

Family transmission barely registers. In the model a child who hears Basque at home starts at 30 points, and school then adds 60. School covers 77% of children whatever their parents speak. The home head start is overwritten before anyone is old enough to count. That is a property of the model rather than a finding about Basque families, and it marks where the model would need extending to study transmission in depth.

Testing language policies

With the model calibrated, each policy can be tested as a what-if: getting more speakers to open in Basque, teaching Basque to adult immigrants, expanding adult classes and changing the school models. Each one was run 200 times in pairs, with the same random seeds with and without the policy, so the comparison isolates the policy from the noise.

Effect by
0+5+10+15+20pts20% more fluent speakers open in BasqueAdult classes reach 1% of people a year, not 0.3%Half of adult immigrants reach a basic levelA quarter of adult immigrants reach a basic levelEvery Spanish-medium (A) pupil moves to Basque-medium (D)A schools teach 3 points a year instead of 1B and D schools teach 8 points a year instead of 6
  • Calibration A: talkative
  • Calibration B: forgiving
  • Whiskers: 95% interval of the mean effect
Percentage points gained in the share of speakers aged 16+, each policy against the same 200 runs without it (same random seeds, so the noise cancels). All policies start in 2006. Both calibrations fit the survey data equally well; where their dots disagree, the data cannot tell us which is right.

Who speaks first is the strongest lever. If 20% more of the fluent speakers opened in Basque, the share of speakers would be 7 to 9 points higher by 2034, close to what tripling adult classes would do. The effect is clear within the first three years and grows roughly in proportion to the push, by 1.3 to 2 points by 2024 for every extra 5% of speakers.

Adult classes add speakers one by one. The gain is linear in the enrolment rate, with no knock-on effect through conversations. Adults leave the euskaltegi at 10 points, in the band where every conversation in Spanish costs them.

Teaching immigrants helps, slowly. The effect grows with time, since each year’s newcomers add to the last. It depends little on the target level: 5, 15 or 30 points give similar gains, because the model counts anyone with some fluency as a speaker.

School reform works through access, not hours. Moving every A-model pupil to D adds 1 to 2 points. Adding hours in models B and D beyond the calibrated 6 points a year adds nothing, because pupils who leave school very fluent are the ones who lose most in conversations in Spanish.

Where it goes next

Agent-based models are only as good as the rules they are given, so the natural next steps are about tightening those rules:

The result that holds across every calibration is the one with the clearest message: school builds speakers, and whether they keep their Basque depends on how often someone dares to speak it first.

Tools. Python (NumPy, SciPy, Matplotlib, multiprocessing), KD-trees for the nearest-neighbour search. The figures on this page come from 200 runs of the model per scenario; the conversation demo runs in your browser.

Data. Basque Government Sociolinguistic Survey (Inkesta Soziolinguistikoa, 2006–2021), Eustat population, migration and school enrolment series, HABE adult-learning indicators.

Credits. Master’s thesis by Maddi Aldaiturriaga Tuñon, supervised by Iñaki Iurrebaso Biteri, 2025–2026. Model design, implementation, calibration and simulations by me. Builds on Civico (2019), The Dynamics of Language Minorities, JASSS 22(4).