Wordfish: parliaments by their words
This is an experiment: it places the members of a parliament on a single axis using nothing but the words they use in their speeches. No votes, no party labels, no hand-coding, just word frequencies. Below is how it works, how to read it, and the parliaments where enough speech text exists to try it.
Parliaments you can explore
Each of these has enough transcript text in the recent window to draw a chart. “Members placed” is how many members the scaling can position; “Speech data” is the full span of speech text available for that parliament.
| Parliament | Members placed | Speech data |
|---|---|---|
| Switzerlandcountry | 252 | 1999–2026 |
| Berncanton | 160 | 2003–2026 |
| Thurgaucanton | 115 | 2008–2026 |
| St. Gallencanton | 90 | 2025–2026 |
| Berncity | 70 | 2007–2026 |
| Basel-Stadtcanton | 67 | 2010–2026 |
| Juracanton | 48 | 2020–2026 |
| Appenzell Innerrhodencanton | 42 | 2025–2026 |
What it does
For a chosen parliament, we take every active member and all of their speeches within a shared recent window. It is the same span for everyone, so the result reflects word choice and not who has been in office longest. Each member becomes one “document”: a bag of the words they used and how often.
A statistical model called Wordfish then finds the single axis of word usage that best separates the members. It discovers which words pull people apart and places each member accordingly. The axis is not defined in advance. It emerges from the speeches themselves.
How it is computed
Every speech is split into words and lowercased, keeping accents. Very common function words (“the”, “and”, “der”, “le” …) are removed with a stopword list matched to each speech's own language, so a French speech is filtered against French stopwords and a German one against German. Rare words that only a handful of members ever use are dropped too, since they add noise, not signal.
Wordfish models each word count as a Poisson draw that depends on the member's position, the word's typical frequency, and how strongly that word discriminates. It fits these in alternating rounds until the numbers stop moving, then standardises the positions so the spread is comparable. The maths runs on demand while the page loads. For a large chamber over a year it takes several seconds, which is why the chart streams in behind a spinner.
How to read the chart
Each dot is a member, placed by their words: members who talk alike sit close, members who talk very differently sit far apart. The direction of the axis is arbitrary. Wordfish finds the split, not which side is “left” or “right”. To see what each end means, read the words listed at the two poles.
The spread is standardised, so distances are a convention rather than a real measure. Each member also carries an uncertainty (the error bar): members who spoke little are placed with low confidence and drawn as hollow rings, so don't over-read them. And speaking style reflects role as much as belief: committee chairs, ministers and frequent speakers can land at a pole because of what they do, not only what they think.
Why most parliaments are empty
The scaling needs the actual transcript text of speeches, and the open data only holds that for a few parliaments. Most bodies have no speeches at all; some have speech records without any transcript; and a few have transcripts that stop years ago. Only the parliaments listed below currently have enough recent speech text for at least ten members to be placed.
References & further reading
- Slapin & Proksch (2008): “A Scaling Model for Estimating Time-Series Party Positions from Texts”, American Journal of Political Science. The paper that introduced Wordfish.
- Wordfish tutorial (quanteda): a hands-on description and reference implementation of the model.
- Grimmer & Stewart (2013): “Text as Data”, Political Analysis. A survey placing Wordfish among text-scaling methods.