For researchers – Sanamahti

Sanamahti is a Finnish-language word game in which players find words on a letter grid by swiping. The game is meant to become an instrument that measures which words its players know and which words they expect Finnish to have. The method has been designed but not locked, and now is the easiest time to influence it.

Status on 12 October 2026. The game exists and is played in the browser and in Android and iOS apps. The measurement system, the weekly lexicon editions and the open data are still at the design stage, and the first experiments have been run. Nothing from them has been published yet.

The game and most of this site are in Finnish. The pages for researchers are also available in English.

What the project is

A player forms words from adjacent letters on the grid and scores points by word length. The game accepts words only in their base form. Every swipe is a small experiment: the player offers a string as a word, and the game answers yes or no.

Today the game’s lexicon is fixed: every board has a precomputed word list that does not change. Some common words are missing from the lexicon, and it contains words that do not belong there. A living lexicon has been designed to replace it. In the design it is updated once a week as a numbered edition:

The week’s rejected swipes, word proposals from players and familiarity ratings are collected.
A rule-based morphological analyser separates inflected forms, derivatives and compounds.
For the remaining candidates, dated evidence is retrieved from outside the game, for example from a news corpus.
An automated AI editor decides within written rules. Each decision is published with its rationale and evidence.
Players can challenge a decision for a week before it takes effect.

Sanamahti is the project of one developer, Alex Liao. There is no research group or dictionary editorial team behind it. That is why the method is written out in the open, and why this invitation exists.

What data is planned

In the design, the system produces three kinds of information. No data fit for research exists yet.

Measured familiarity with pseudo-word controls

Familiarity is to be measured like this: a player gets a strip of ten words and says for each one whether they know it. Three of the ten are pseudo-words. They are used to correct the responses for guessing and for a tendency to answer yes. The result is one figure per word: the share of respondents who know it.

Nykysuomen sanalista, the word list of contemporary Finnish published by the Institute for the Languages of Finland (Kotus), records up to four things about a word: a headword, a homonym number, a word class and an inflection type. It does not say how many people know the word.

Words that players expect but the lexicon lacks

A rejected swipe is a string that a player offered as a word and the lexicon did not recognise. In the design, the analyser classifies each one: an inflected form of a word in the lexicon, a derivative, a compound or unknown. The data is meant to show which inflected forms players offer as base forms and which words they expect to find in the lexicon.

Automated editorial decisions that can be checked

The background research for the project found no dictionary or game lexicon that is updated without human approval or a reference source. In Sanamahti’s design the decisions are made by AI, so every decision has to be open to checking.

For that reason a case is opened for each candidate word. The case stores the evidence with its sources and dates, the decision, the rationale and references to the rules the decision relies on. The decision log is intended to be published. It will then be possible to study where automated editing succeeds and where it goes wrong.

What the data is not

The players are not a sample of Finnish speakers. The composition of the player base has not been measured, and no background information is attached to the observations. A swipe is a hint, not a vote for a word, and the grid limits what can be swiped. The limitations are described on the method page.

What collaboration could look like

Reviewing the method. Read the method description and point out where it is weak. The thresholds, the guessing correction formula and the way pseudo-words are generated are still open.
The word jury. The rules of the lexicon are to be derived from ratings that invited raters give to word cards: is the word acceptable in the game or not, and why. The plan is to invite language researchers, teachers of Finnish and journalists, among others. If you would like to take part, say so through the contact form.
Your own research question. If your research needs a measurement that the design does not include, say so. The structure is still open.
Theses. There are six topics for bachelor’s and master’s theses. They can be offered to students as they are or adapted.
Using the data. The data is intended to be published as open and citable. The plan and the state of the licence are on the data page.

Get in touch

Write through the contact form and choose researcher as your role. Say briefly what interests you in the project, or what you doubt. The messages are read by Alex Liao.