统计机器翻译中大规模特征的深度融合

Deep combination of large-scale features in statistical machine translation

对循环神经网络和递归神经网络进行改进,提出深度融合的神经网络(DNN)模型,在训练过程中加入大规模特征.该模型有很强的泛化能力,适合于现在主流的自底向上解码样式,融合了2种经典的机器翻译模型:基于短语的层次化文法(HPG)和括号转录文法(BTG).使用改进的循环神经网络,生成适合短语生成过程的短语/规则对语义向量,并在生成过程中使用了自编码器以提高循环神经网络的性能.使用改进的递归神经网络,使它在翻译过程中指导解码,考虑到另一个解码器在解码过程中的信息,互相影响共同提高翻译性能.提出的深度融合模型不仅适合于异类翻译系统,也适合于异类语料.相对于经典的基线系统,在异类系统上该模型的实验结果获得1.0~1.9倍的BLEU分数提高,在异类语料上该模型的实验结果获得1.05~1.58的BLEU分数提高,且进行了统计显著性检验.

Deep neural network（DNN）has many successful applications in statistical machine translation（SMT）,and the absent semantic problem of machine translation system was solved.The mainstream recurrent neural network（RTNN）and recursive neural network（RENN）model were modified,and a deep neural network combination（DCNN）of large-scale features for system combination in SMT was presented.The model has strong generalization ability,which is suitable for the current mainstream bottom-up decoding style. Hierarchical phrase-based grammar（HPG） was combined with bracket transduction grammar（BTG）.The improved recurrent neural network was used to generate the phrasepair semantic vector which is suitable to phrase generation process,and the autoencoder was used to improve the performance of the recurrent neural network.The improved recursive neural network was used to guide the decoding process in SMT task,and the mutual influence information was considered from another decoder.The deep neural translation combination model is suitable not only for heterogeneoussystem,but also for heterogeneous corpus.The experimental results showed that DCNN significantly improved the performance of a state-of-the-art SMT baseline system,leading to a gain of 1.0-1.9and1.05-1.58 BLEU points in heterogeneous system and corpus combination,respectively.