Write letters in the well-known and beloved quality and "Herzensgüte" we are used to from letters of official German institutions.
What it does
Generate random letters similar in style to letters in "Verwaltungssprache" (German officialdom).
How we built it
We used textgenrnn as neural network and trained it on press releases to gain a solid German language model. Afterwards, we trained the model on letters in "Verwaltungssprache" that were gained by optical character recognition from scanned documents. We built a web page that displays the generated text in a letter format.
Challenges we ran into
Quality of the available data. Runtime of the training process on datasets of even average size, due to laptop hardware.
Accomplishments that we're proud of
We got it working! We have successfully got first results of generated text. Also our frontend slowly coming together. Isn't that really cool? We are proud of that.
What we learned
Enough for two Hackathons. Machine learning was new for some project members, as was HTML & CSS.
What's next for S.T.U.M.P.F.
Write more letters. And hopefully gain some pan pals.