Monotonic filter for hierarchical translation models

Abstract

The model size and decoding time are known issues in statistical machine translation. Especially, monotonic words order of language pairs makes the size of hierarchical models huge. Considering this fact, the rule extraction method of phrase-boundary model was changed to extract less number of rules. This paper proposes this rule extraction method as a general filter for hierarchical models. Named as monotonic filter, this filter reduces the extracted rules from phrase pairs decomposable to monotonic aligned subphrases. We apply the monotonic filter on the hierarchical phrase-based, SAMT and phrase-boundary models. Our experiments are performed in translations from Persian, German and French to English as the source and target languages with low, medium and high monotonic word order respectively. The reduction amount of the monotonic filter for the model size and decoding time is up to about 70% and 80% respectively, in most cases with no tangible impact on the translation quality.