Statistical Machine Translation for Twitter

We consider the problem of translating short messages (Tweets) using Europarl as a starting-point. After highlighting some of the domain differences between Europarl and Twitter, we show that for German-English translation, we can improve performance from a baseline BLEU score of 25.58 to 53.45. By far and away the single most important improvement is passing-through unknown words (which are mainly URLs). Enforcing the length constraint upon translated output turnsout to be relatively simple. Since our Twitter translation involves little reordering, we conclude that the biggest challenge is lexical: dealing with unknown words, spelling mistakes, creative orthography and Twitter-idioms.