The Universe of Discourse

Mon, 12 May 2008

Artificial Finnish

By 1988 or 1989 I had read in several places, most recently in J. R. Pierce's Symbols, Signals, and Noise, that if you compile a table of the relative frequencies of three-letter sequences (trigraphs) in English text, and then generate random text with the same trigraph frequencies, the result cannot be distinguished from meaningful English text except by people who actually know English. Examples were provided, containing weird but legitimate-sounding words like "deamy" and "grocid", and the claim seemed plausible. But since I did actually know English, I could not properly evaluate it.

But around that time the Internet was just beginning to get into full swing. The Finnish government was investing a lot of money in networking infrastructure, and a lot of people in Finland were starting to appear on the Internet.

I have a funny story about that: Around the same time, a colleague named Marc Edgar approached me in the computer lab to ask if I knew of any Internet-based medium he could use to chat with his friend at the University of Oulu. I thought at first that he was putting me on (and maybe he was) because in 1989 the University of Oulu was just about the only place in the world where a large number of people were accessible via internet chat, IRC having been invented there the previous autumn.

A new set of Finnish-language newsgroups had recently appeared on Usenet, and people posted to them in Finnish. So I had access to an unlimited supply of computer-readable Finnish text, something which would have been unthinkable a few years before, and I could do the experiment in Finnish.

I wrote up the program, which is not at all difficult, gathered Finnish news articles, and produced the following sample:

Uttavalon estaa ain pahalukselle? Min omatunu selle menneet hy, toista. Palveljen alh tkö an välin oli ei alkohol pisten jol elenin. Että, ille, ittavaikki oli nim tor taisuuristä usein an sie a in sittä asia krista sillo si mien loinullun, herror os; riitä heitä suurinteen palve in kuk usemma. Tomalle, äs nto tai sattia yksin taisiä isiäk isuuri illää hetorista. Varsi kaikenlaineet ja pu distoja paikelmai en tulissa sai itsi mielim ssän jon sn ässäksi; yksen kos oihin! Jehovat oli kukahdol ten on teistä vak kkiasian aa itse ee eik tse sani olin mutta todistanut t llisivat oisessa sittä on raaj a vaisen opinen. Ihmisillee stajan opea tajat ja jumalang, sitten per sa ollut aantutta että voinen opeten. Ettuj, jon käs iv telijoitalikantaminun hä seen jälki yl nilla, kkeen, vaaraajil tuneitteistamaan same?

In those days, the world was 7-bit, and Finnish text was posted in a Finnish national variant of ASCII that caused words like "tkö an välin" to look like "tk| an v{lin". The presence of the curly braces heightened the apparent similarity, because that was all you could see at first glance.

At the time I was pleased, but now I think I see some defects. There are some vowelless words, such as "sn" and "t", which I think doesn't happen in Finnish. Some other words look defective: "ssän" and "kkeen", for example. Also, my input sample wasn't big enough, so once the program generated "alk" it was stuck doing the rest of "alkohol". Still, I think this could pass for Finnish if the reader wasn't paying much attention. I was satisfied with the results of the experiment, and was willing to believe that randomly-contructed English really did look enough like English to fool a non-English-speaking observer.

[ Addendum 20080514: There is a followup to this article. ]

[ Addendum 20080601: Some additional notes. ]

[Other articles in category /lang] permanent link