深度递归的层次化机器翻译模型

发布时间：2018-04-30 13:22

本文选题：循环神经网络 + 递归神经网络　；参考：《计算机学报》2017年04期

【摘要】：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1.49~1.84BLEU分数.
[Abstract]:Deep learning has many applications in natural language processing. The main function of the depth network is to capture the deeper semantic information hidden in the language structure. This paper proposes a hierarchical machine translation model with depth recursion, which is guided by the alignment in the original sentence as the generation structure of the depth network, and combines the advantages of the original depth translation model. Compared with the existing neural translation models, the hierarchical translation process is better combined with the advantages of cyclic neural networks and recurrent neural networks. The induction of hierarchical rules consists of two parts: inductive of phrases and induction of formal rules. In the modeling process of this paper, the two parts are simulated and accord with the inductive process. In this paper, word-level semantic errors, monolingual phrase / rule semantic errors and bilingual phrase / rule semantic errors are used to construct objective functions in the training. At the same time, the alignment information is taken into account to guide the training of hierarchical depth neural networks. In the process of decoding, the semantic relationship between sentences can be obtained by generating the semantic vector of some translation results, which can add semantic information to the grammatical structure and overcome the lack of semantic information in the original hierarchical model. The experimental results show that the hierarchical machine translation model with deep recursion is effective and improves the 1.49~1.84BLEU score compared with the classical baseline system.
【作者单位】：哈尔滨工程大学计算机科学与技术学院;哈尔滨理工大学软件学院;
【基金】：国家自然科学基金(61300115) 中国博士后科学基金(2014M561331) 黑龙江省教育厅科技研究项目(12521073)资助~~
【分类号】：TP183;TP391.2

【相似文献】