I recently updated my @ZwoSchlagzeilen Twitter bot. I replaced the complicated, mostly rule-based language generation algorithm with a statistical approach using a trigram language model. Details can be found in the GitHub repository.

Using this approach, the bot strictly speaking doesn’t mix two headlines anymore but rather uses a part from a randomly sampled headline as “seed” headline and then randomly samples new words conditional on the previous two words. So it mixes one headline with words from a corpus of headlines (originating from news of the last two weeks).

I observed the generated tweets for the last two weeks and found the results better than those generated with the previous approach. Still, some generated headlines don’t make sense at all, but these imperfections are what makes it fun.