Replacing Linguists with Dummies: A Serious Need for Trivial Baselinesin Multi-Task Neural Machine Translation

Recent developments in machine translation experiment with the idea that a model can improve the translation quality by performing multiple tasks, e.g., translating from source to target and also labeling each source word with syntactic information. The intuition is that the network would generalize knowledge over the multiple tasks, improving the translation performance, especially in low resource conditions. We devised an experiment that casts doubt on this intuition. We perform similar experiments in both multi-decoder and interleaving setups that label each target word either with a syntactic tag or a completely random tag. Surprisingly, we show that the model performs nearly as well on uncorrelated random tags as on true syntactic tags. We hint some possible explanations of this behavior. The main message from our article is that experimental results with deep neural networks should always be complemented with trivial baselines to document that the observed gain is not due to some unrelated properties of the system or training effects. True confidence in where the gains come from will probably remain problematic anyway