Neural network-based reranking model for statistical machine translation

The non-local feature always plays an important role in improving performance of SMT. Nonlinear neural network model can take better advantage of non-local features to improve the performance of translation through the introduction of the hidden layer. So this paper will build reranking models based on neural network to make use of non-local features to improve the translation performance. In this paper, we will introduce two models: Reranker-WC and Reranker-D. Compared with performance of the baseline system, the performance of Reranker-WC can be promoted to about 1.4 BLEU score. Moreover, we find that different hyper-parameter λ will also affect the quality of SMT output at the same time. We achieve the best performance while λ is 40.