Deep Learning has many applications in natural language processing.The main role is to capture the deeper semantic information hidden in the language structure through the deep network.The motivation of this paper is that we use the word alignment of the bilingual sentence as the guide to generate the structure of deep network,and combine these advantages of the original deep translation model.The paper proposes Hierarchical Recursive Neural Network (HRNN) for hierarchical machine translation model.Compared with the existing neural translation model,the model is a better combination of phrase-based hierarchical translation model and deep neural network.It has two advantages of Recurrent Neural Network (RTNN) and Recursive Neural Network (RENN).The procedure of phrase and formal rule induction can be simulated in HRNN,and the model meets induction procedure.In training procedure,the objection function of this paper include the monolingual word-level semantic errors,the monolingual phrase/rule semantic errors and the bilingual phrase/rule semantic error,and the semantic effect of three parts are balanced in statistical machine translation (SMT).In decoding procedure,the semantic relation among sentences are obtained by the semantic vector of partial translation result,and this method,which the semantic information is added to the syntax structure,overcomes the lack of semantic information in the original model.The experimental results show that HRNN significantly improves the performance of a state-of-the-art SMT baseline system,leading to a gain of 1.49~1.84 BLEU points.%深度学习在自然语言处理中有很多的应用.深度网络的主要作用是捕获隐藏在语言结构中更深的语义信息.该文出发点为根据原有句子中的对齐作为深度网络生成结构的指导,并融合原有深度翻译模型的优点,提出了深度递归的层次化机器翻译模型.相对于已有的神经翻译模型来说,更好地结合了层次化的翻译过程,同时这种方法结合循环神经网络和递归神经网络的优点.层次化规则的归纳包含两个部分:短语的归纳和形式化规则的归纳,而在该文的建模过程中模拟了这两个部分且符合归纳过程.该文在训练中采用单词级语义错误、单语短语/规则语义错误和双语短语/规则语义错误构造目标函数,训练中能够更好平衡语义中3个部分的影响,同时考虑到对齐信息以指导层次化深度神经网络的训练.在解码过程中通过生成部分翻译结果的语义向量,最终得到句子间的语义关系,这样可以在语法结构中加入语义信息,克服了原有层次化模型语义信息缺乏的问题.该模型的实验结果说明了深度递归的层次化机器翻译模型的有效性,相对于经典的基线系统提高了1.49~1.84 BLEU分数.
展开▼
