Improving Relative-Entropy Pruning using Statistical Significance

Relative Entropy-based pruning has been shown to be efficient for pruning language models for more than a decade ago. Recently, this method has been applied to Phrase-based Machine Translation, and results suggest that this method is comparable the state-of-art pruning method based on significance tests. In this work, we show that these 2 methods are effective in pruning different types of phrase pairs. On one hand, relative entropy pruning searches for phrase pairs that can be composed using smaller constituents with a small or no loss in probability. On the other hand, significance pruning removes phrase pairs that are likely to be spurious. Then, we show that these methods can be combined in order to produce better results, over both metrics when used individually.