Abstract
Recent studies have revealed a number of pathologies of neural machine translation (NMT) systems. Hypotheses explaining these mostly suggest that there is something fundamentally wrong with NMT as a model or its training algorithm, maximum likelihood estimation (MLE). Most of this evidence was gathered using maximum a posteriori (MAP) decoding, a decision rule aimed at identifying the highest-scoring translation, i.e. the mode, under the model distribution. We argue that the evidence corroborates the inadequacy of MAP decoding more than casts doubt on the model and its training algorithm. In this work, we criticise NMT models probabilistically showing that stochastic samples following the model's own generative story do reproduce various statistics of the training data well, but that it is beam search that strays from such statistics. We show that some of the known pathologies of NMT are due to MAP decoding and not to NMT's statistical assumptions nor MLE. In particular, we show that the most likely translations under the model accumulate so little probability mass that the mode can be considered essentially arbitrary. We therefore advocate for the use of decision rules that take into account statistics gathered from the model distribution holistically. As a proof of concept we show that a straightforward implementation of minimum Bayes risk decoding gives good results outperforming beam search using as little as 30 samples, confirming that MLE-trained NMT models do capture important aspects of translation well in expectation.
Abstract (translated)
URL
https://arxiv.org/abs/2005.10283
