How far apart are Finnish parties?
What 130,000 parliamentary speeches reveal about polarisation, 2015–2026
At 7.57am on a May morning in 2021, the Speaker’s gavel finally fell on the longest sitting of the decade. Parliament had argued about the EU recovery package all night, and 99% of the small-hours speeches came from a single party. Interjections shouted across the chamber have doubled since 2015. The word polarisaatio is uttered three times as often as a decade ago.
It certainly feels like Finnish politics is pulling apart. But feelings are cheap. So we measured it: we taught a computer to guess which party a speaker belongs to from a single plenary speech — no names, no party labels, just the words — and watched how much easier that guess has become over three parliamentary terms.
130,000 plenary speeches, 2015–2026
1 in 3 correct party guesses today, up from 1 in 4
0.39 → 0.56 separability of party pairs
0.13 / 0.76 closest and furthest party pair today
The step came in 2023
Our headline measure is separability: how reliably two parties’ speeches can be told apart (0 = indistinguishable, 1 = always distinguishable). A separability of 0.56 means that, shown one speech from each of two parties, the model tells them apart correctly about 78 times in 100; a coin flip would score 50. Averaged over all 28 party pairs, the measure climbed from 0.39 to 0.45 to 0.56 across the three terms. That is not a gentle drift. Most of the move happens in one step.
Why believe a computer’s guessing game? Two checks. Shuffle the party labels and repeat everything, and the measure collapses to zero: the method invents nothing. And a completely different yardstick from the economics literature, the leave-out estimator that Gentzkow et al. (2019) built for the US Congress and Simola et al. (2023) applied to a century of Finnish parliamentary speech, ranks the party pairs almost identically.
That second yardstick can be computed year by year, and it sharpens the story. From 2015 to 2022 the line wobbles without a sustained trend: it drifts up in 2018–21, then falls back. From 2023 it climbs above every earlier year and stays there. What looks like a decade-long trend is closer to a step change during the current term, with one caveat: the partial year 2026 dips back towards the old range. Part of the earlier rise even has a mundane explanation: speeches simply got longer, and longer speeches are easier to classify. The jump after 2023 survives that correction.
It is not just that parties talk about different things
An obvious objection: perhaps parties have not changed how they speak, only what they speak about. The Finns Party talks about immigration, the Greens about climate, and a word counter tells them apart without any polarisation being involved.
So we repeated the measurement inside topics, using the topic keywords of a recent study of Finnish social media (Gronow and Malkamäki 2024). The objection fails. Within every topic — climate, immigration, security, even the tax-heavy inequality debate — parties have become easier to tell apart.
Parties are not merely choosing different subjects. They increasingly speak about the same subject in different ways. The timing matches the Twitter study, too: the gap between the Finns Party and the left-green parties on climate doubled in 2019–23, exactly when climate began to polarise on Finnish social media. Part of the within-topic gap is institutional role, since governments defend bills and oppositions attack them in the same debate.
Two blocs
Which parties actually sound alike? The map below places parties so that the ones our model confuses most often sit closest together. The buttons animate the changes from term to term.
The map has no fixed axes — distance is the only thing it shows — but by 2023–27 its layout happens to reproduce the government–opposition divide. Three things stand out.
The right-side trio sounds alike everywhere. NCP Finns CD speak alike in every term, including 2019–23, when all three sat in opposition. Their speech is more similar than their election-compass answers would predict. Government status alone does not explain it: the five-party Marin government spoke less cohesively than its three-party opposition.
Close in views, far in voice. The Left Alliance, the Social Democrats and the Greens — Left SDP Greens — are among the closest pairs on the election compass, yet each speaks in its own register.
In 2023–27 the two dividing lines coincide for the first time. The right-side trio is now the core of the government, the left-green parties lead the opposition, and the pairs across that line have become the most separable pairs in parliament: 0.66 on average, against 0.45 two terms ago. That coincidence is what produces the step in the first figure. The extremes put numbers on it: the single closest pair in today’s parliament is the Christian Democrats and the National Coalition (0.13), while the Finns Party against the Left and against the Greens (0.76 each) are the furthest apart, once the Swedish People’s Party’s topic-driven outlier is set aside.
And the lone dot at the edge of every map? The Swedish People’s Party is separated by its subjects — education, the Swedish language, its regions — rather than by conflict with anyone. The map measures speech, not positions.
The data also rule something out: the parties pulled apart, but their MPs did not fall into line. The spread of individual MPs around their own party’s centre has not narrowed: parliament has not filled with clones. The distance grew between the parties, not by discipline within them.
The words that give each party away
What does the model actually latch onto? Its most party-revealing words are no mystery to anyone who follows Finnish politics, and that is the point: the machine rediscovers each party’s vocabulary from scratch. A sample from the current term (Finnish word stems, translations in italics):
Finns maahanmuutto immigration rikos crime veronmaksaja taxpayer
Left työntekijä worker pienituloinen low-income leikkaus cut
SDP palkansaaja wage earner työelämä working life
Greens ilmasto climate luonto nature päästöt emissions
Centre metsä forest maatalous agriculture Itä-Suomi Eastern Finland
NCP talous economy työ work uudistus reform
Role shows up too: whoever sits in opposition is recognisable by how much they talk about the government and the prime minister.
NoteHow we measured this
We trained a classifier (TF-IDF features and logistic regression) to predict the speaker’s party from a single speech, separately for each term and eight parties, from about 130,000 plenary speeches (2015 – September 2026, Parliament of Finland open data). Names, party names and formulaic phrases are removed before training. The model is always tested on speakers it never saw during training, every party gets the same amount of training data, and every number is reported against a shuffled-labels baseline, the precautions that Gentzkow et al. (2019) show are necessary before reading classifier accuracy as polarisation. Separability of a party pair is 2·AUC − 1 on the model’s probabilities; our cross-check is the leave-out estimator used by Simola et al. (2023). The measure also passes an external check: party distances in speech correlate with the parties’ distances on Yle’s election-compass answers in the two most recent terms (Spearman 0.6–0.7 without the Swedish People’s Party). Full details, robustness checks and code: the technical report and repository.
A century of perspective
Is this the most divided Finnish parliament has ever been? Almost certainly not. Simola et al. (2023) ran essentially the same measurement over every plenary speech since 1907, and their century-long series peaks in the 1970s, driven by the pro-Soviet far left. By that measure, Finnish parliamentary speech has been moderate for most of its history, far below the levels measured in the US Congress. Partisanship has crept up since the 1990s, but gently.
Their series stops in 2018, one term into our window. Our data pick up the story from there and find the clearest shift only after 2023. The two studies use different measures, so their levels cannot be laid on the same axis, but read together they suggest that whatever is happening in today’s chamber, it is recent, and it was not yet visible when the century-long view was last taken.
Side notes from the transcripts
The transcripts record more than words: gavel knocks, laughter, interjections. A few favourites:
Parliament heckles along bloc lines. Interjections have doubled, from about 28 per hundred speeches in 2015–19 to 56 now, and their direction follows the blocs: left-green MPs heckle the National Coalition and the Finns Party, and vice versa.
The chamber has noticed its own drift. Words for polarisation itself (polarisaatio, polarisoitua and their relatives) appear three times as often in the current term (22 occurrences per million words) as in 2015–19, and the number of different MPs using them has grown at the same rate.
Every year has its word. The fastest-rising stems trace the decade:
2016 kiky 2017 aktiivimalli 2018 osakesäästötili 2019 terapiatakuu 2020 korona 2021 koronapassi 2022 Ukraina 2023 indeksijäädytys 2024 sovittelija 2025 velkajarru
A party split you can hear. When the Blue Reform group broke away from the Finns Party in 2017, its speech went with it: by language the breakaways sat as close to their coalition partner, the National Coalition (0.19), as to the party they had just left (0.21).
And the longest night belongs to the EU recovery package: the sitting of 12 May 2021 ended at 7.57am the next morning, and 99% of the small-hours speeches came from one party: the Finns Party.
What this does and does not show
Three caveats belong in any honest reading. First, three terms means three different governments: with these data, a time trend cannot be fully separated from who happens to govern. Second, we measure language, not opinions, though the speech distances do match election-compass distances in the two most recent terms. Third, the smallest parties (the Christian Democrats and the Swedish People’s Party) have so few MPs that their numbers are noisy, and the current term is still incomplete (data to September 2026). Unlike Simola et al. (2023), we do not control for MPs’ region or gender.
What the data do show is specific: a measurable, robust widening of the linguistic distance between Finnish parties, concentrated in the current term, visible within topics, and organised around two blocs whose boundary now coincides with the government–opposition divide.
Data: Parliament of Finland open data (CC BY 4.0) and Yle election compasses (CC BY 4.0), analysed at party level only. AI (Anthropic’s Claude) was used for research design, code, figures and text drafting; the team reviewed all output. Code and methods: github.com/jtompuri/data-science-mini-project.
References
Gentzkow, Matthew, Jesse M. Shapiro, and Matt Taddy. 2019. “Measuring Group Differences in High-Dimensional Choices: Method and Application to Congressional Speech.” Econometrica 87 (4): 1307–40.
Gronow, Antti, and Arttu Malkamäki. 2024. Political Polarisation in Turbulent Times: Tracking Polarisation on Finnish Social Media 2015–2023. Helsingin Sanomat Foundation.
Simola, Salla, Jeremias Nieminen, and Janne Tukiainen. 2023. “A Century of Partisanship in Finnish Political Speech.” Aboa Centre for Economics Discussion Paper, no. 160.