Revisiting Multi-Domain Machine Translation

When building machine translation systems, one often needs to make the best out of heterogeneous sets parallel data in training, and robustly handle inputs from unexpected domains testing. This multi-domain scenario has attracted a lot recent work that fall under general umbrella transfer learning. In this study, we revisit translation, with aim formulate motivations for developing such systems associated expectations respect performance. Our experiments large sample show most these are hardly met suggest further is needed better analyze current behaviour them fully hold their promises.