Last year I was invited to interview at Google Aarhus for a job on exactly this topic. We discussed fun, very simple ways to avoid including loads unused library code, and I was told that they already had some ridiculously efficient, practical way of compressing translation strings. I was there last November, and I had the impression that a lot of low hanging fruit would be picked off very quickly. Not sure what parts of this work actually shipped in these last 11 months. But work may still be ongoing.