Stochastic Report Generator • Home

Forget about LLMs or even Markov Chains, sometimes correct grammar is all you need for perfectly correct gibberish: I'm happy to annnouce my translation of the stochastic report generator(stokastiske rapportgenerator).

The story of the generator goes back to 1996 where it was released by Peter Sestoft, now head of the ITU Computer Science Department. There is a good write up (in danish) about it here, but the jist of is that he was tired of the then trend in the education sector of creating "virtual centers". He had, while in high school, written some software for producing grammatically correct danish nonsense, and now extended this into a web-available generator of proposals of new "virtual centers".

I first encountered it in about 2012 while studying at ITU and thought it was hilarious, but I was also fascinated with how a program like that would work. It was quiet different from the kinds of program I had previously been exposed to. So at some point I decided to try and read the source code and port the program from MoscowML to F#, which I was learning at the time. I made some progress, but eventually hit a wall and gave up.

Recently however, the generator went offline and I decided it was time to revisit the project, so here we are. There are still some bugs and missing features, but the core is there, ported to F# and hosted online as an Azure Function.

What did I learn?

The basic type it works on is ordsek (word sequence) One interesting aspect of the program is that it makes extensive use of custom operators.

&& Is used as an infix operator to concatenate two word sequences. &&& Is also an infix operator, but this one evaluates two functions and concatenates the results.

Because F# already uses && for boolean AND. I translated these as &&& and @@@ respectively. It also uses a few custom operators for random choices. I really like the >>> operator:

  0.25 >>> something

Lets you do something, but only 25% of the time.

And also || which is an infix operator that lets you pick one of two options, so you can write:

val text =
    intro
    &&& discussion
    &&& (0.25 >>> (figureA || figure B))
    &&& conclusion

Notice that intro etc. here are function calls and not previously evaluated. Neat.

Saxo Grammaticus

The most important element is of course how to handle the correct construction of sentences. We have already seen the structure of a text in sections. Each section, intro, consists of a set of sentences. Theses in turn are picked from a small set of options, like this one:

let ledsaetning = nominal @@@ adverbial @@@ verbPraesIndAkt @@@ nominal

The first step here is to pick a noun (nominal). The avialable nouns are hardcoded into the sourcecode like this:

 RegS (Fk,  t, "prototype", "-n", "-r", "-ne", [|""; "software"|]);
 RegS (Itk, t, "system", "-et", "-er", "-ne", 
       [|""; "edb-"; "informations"; "IT-"; "kommunikations"|]);

This includes enough information to pick a random form of the noun, including a random prefix, like "IT-Systemet". The we pick a random adverb, again from a hardcoded list, like ""i ringe grad" Then, I guess, a verb in present indicative active form, from another list, like "komplicerer". And finaly anoter nominal, say "prototypen".

This type of sentence is called after a conjunction, like "da".

We get: "Da IT-systemet i ringe grad komplicerer prototypen." loosely translated: "Since the IT-System almost does not complicate the prototype."

Tada!