PARAMETER ADJUSTMENT METHOD USED FOR STATISTICAL MACHINE TRANSLATION

摘要

The present invention relates to a parameter adjustment method. The parameter adjustment method used for statistical machine translation comprises: step 1: using a monolingual corpus to construct a language model required for translation; step 2: using a bilingual parallel corpus to construct a translation model; and step 3: using a target function to adjust a parameter. The method can solve problems in prior art that the parameter is easily over-fit or falls into a local optimum during a parameter adjustment process, and the method is easy to implement and can combine a larger number of features. In addition, the target function is convex; therefore, global optimum can be achieved in a training process.

主权项

1. A parameters adjustment method in a statistical machine translation, characterized in that, the method comprises the following steps:
Step 1: building a language model required for translation by utilizing a monolingual corpora; Step 2: building a phrase translation model by utilizing a bilingual parallel corpus; Step 3: Step 3: processing parameters adjustment for λm by utilizing an objective functionminλ∑s=1n[-Σm=1Mλmhm(es,fs)+log∑e∈Csexp{Σm=1Mλmhm(fs,e′)+l(e′,es)}],where es refers to reference translation, e′ refers to machine translation, fs refers to sentence in source language awaiting for translation processing, hm (es,fs) and hm (fs,e′) refer to the characteristics used in building the translation system, the characteristics comprises four main categories, which are language model, phrase translation listing, sequence model and correctional word penalty items, m=1, . . . , M, M refers to the total number of characteristics, l(e′, es) refers to cost function, C5 refers to the collection set of machine translation candidate, e′ ε C5.