In this paper, we propose to use various global features for discriminative reranking in an SMT framework. We employ an online large-margin based training algorithm for the structural output support vector machines based on the margin infused relaxed algorithm. Besides the standard features used, such as decoder's scores, source and target sentences, alignments and part-of-speech tags, we include sentence type probabilities, posterior probabilities and back translation features for reranking. These features have been proved to be useful in other approaches in statistical machine translation but it is the first attempt to apply them in reranking. Our experimental results using 160K BTEC corpus show an improvement of 1–4 BLEU percentage points on Japanese/Chinese to English translation.
展开▼
