XJIPC OpenIR  > 多语种信息技术研究室
深度神经网络在维汉机器翻译中的应用研究
孔金英
学位类型博士
导师李晓
2017-05-21
学位授予单位中国科学院大学
学位授予地点北京
学位专业计算机应用技术
关键词维汉机器翻译 深度神经网络 调序模型 未登录词 神经机器翻译
摘要

以深度神经网络为代表的新一代人工智能技术,已经为图像处理和语音识别等领域带来了颠覆性的变革。作为人工智能领域最为火热的研究方向之一,机器翻译迎来了前所未有的发展契机。工业界和学术界都认为,深度神经网络技术能够帮助机器翻译取得更好的结果。神经网络技术在机器翻译上的应用分为两种类型:一种是利用神经网络技术优化改进传统的机器翻译系统,另一种则是拥有全新架构的神经机器翻译系统。传统的统计机器翻译模型由语言模型、翻译模型、调序模型等组成,前一种方法利用神经网络技术优化这些模块,后一种方法直接通过深度神经网络编码输入的句子信息并解码生成译文。维汉机器翻译作为机器翻译领域的一个分支,有着语言跨度大、语言形态信息丰富等特点。由于维吾尔语句法分析、词法分析等语言学研究的滞后,难以通过加入语言学特征改善基于统计的维汉机器翻译模型性能,影响了维汉统计机器翻译系统译文质量的提升。基于深度神经网络技术,本文针对维汉机器翻译所面临的译文语序问题和未登录词问题,提出了以下解决方案:1. 基于深度学习的调序规则表优化模型针对维汉机器翻译中的译文语序问题,本文提出了一种基于深度学习的调序规则表优化模型。该模型由生成模块,判别模块和基于最小差异的过滤策略组成。生成模块利用递归自动编码机向量化调序规则,判别模块使用多层感知机对调序规则进行打分评价,基于最小差异的过滤策略针对调序规则表进行优化过滤。使用本模型优化后的调序规则表重新训练调序模型用于解码,可以加快维汉机器翻译最终的解码速度和提升最终的译文质量。2. 基于注意力的维汉神经机器翻译模型针对维汉机器翻译中的长距离调序问题,本文提出了一种基于注意力的维汉神经机器翻译模型。该模型在一般的基于注意力的神经机器翻译的基础上进行改进,引进了外界的先验知识。基于注意力的维汉神经机器翻译模型在解码输出译文阶段,综合引入的外界汉语语言模型得分和神经机器翻译模型得分选出最优的候选译文。本模型有效的解决了维汉机器翻译中的长距离调序问题,适用于非规范性文本的翻译,比如口语翻译。3. 基于指针神经网络的维汉机器翻译框架针对维汉机器翻译中的未登录词问题,本文提出了一种基于指针神经网络的维汉机器翻译框架。该框架由前处理模块,改进的基于指针的神经机器翻译模型和后处理模块组成。前处理模块规则化维吾尔语,使用修改过的语料训练一个改进的基于指针的神经机器翻译模型和一个短语翻译模型。改进的基于指针的神经机器翻译模型在转换网络的选择策略上偏向于指针网络,这使得指针网络的功能更加强大。后处理模块首先使用训练好的神经机器翻译模型将维吾尔语句子翻译成粗糙的译文,然后用训练好的短语翻译模型将粗糙的译文重译为最终的译文。该框架可以有效的解决维汉机器翻译中的未登录词问题,并改善译文质量。

其他摘要

The new generation of artificial intelligence technology, take deep neural network(DNN) as representation, has brought breakdown in field of image processing and speech recognition. As one of the most popular research directions in artificial intelligence, Machine Translation now faces an unprecedented opportunity for development. Both industry and academia are confident to achieve better translations on Machine Translation by deep neural network technology.Applications of neural network technology on Machine Translation can be divided into two types: first type utilizes neural network to optimize traditional machine translation model, the other is neural machine translation of a novel architecture. Traditional statistical machine translation model consists of Language model, translation model, reordering model and so on. The former type utilizes DNN to optimize these models respectively, the later type directly encode the input sentence information and generate translation through depth neural network.As a branch of Machine Translation, Uyghur-Chinese Machine Translation is of rich morphology information with language pair involving big difference in syntax. Due to lack of analysis tools in Uyghur syntax, Uyghur lexical and other Uyghur linguistic information, Uyghur-Chinese statistical Machine Translation model is difficult to add Uyghur linguistic features, which handicap the improvement of statistical Machine Translation.Based on the technology of deep neural network, this paper puts forward the following solutions to cope with reordering problem and out-of-vocabulary problem faced by Uyghur-Chinese Machine Translation:1.Filtering reordering table using a novel Recursive Autoencoder model To cope with the reordering problem in Uyghur-Chinese MT, this paper proposes a Deep neural network (DNN) model to prune reordering-table. We cast the task of filtering reordering model as a deep learning problem where our model consists of three parts: a generative module to implement rule embedding, a discriminative module to estimate the quality of reordering rules, a filtering strategy based on minimum difference is designed to filter reordering rule. Using the filtered reordering table by our model to retrain reordering model can accelerate the decoding speed and enhance the quality of final translation. 2. An attention-based Uyghur-Chinese neural Machine Translation model To cope with long-distance reordering problem in Uyghur-Chinese MT, this paper presents an attention-based Uyghur-Chinese neural Machine Translation model introducing extra knowledge: first, we take segmentation and spelling check in Uyghur language, then use the Chinese language corpus to train language model. Next, we use the processed Uygur-Chinese corpus to train an attention-based neural machine translation model. In decoding, our model use a comprehensive score which consist of the score from Chinese language model and the score from neural machine translation model to selected the best candidate translations. Our model can effectively solve the long-distance reordering problem in Machine Translation, is applicable to nonstandard texts translation, such as oral translation.3. A pointer-based Uyghur-Chinese neural Machine Translation frameworkTo cope with the problem of out-of-vocabulary(OOV) in Uyghur-Chinese MT, this paper presents a pointer-based Uyghur-Chinese neural Machine Translation framework. This framework consists of pre-process, modified pointer-based NMT and post-process. Pre-process modify the Uyghur-Chinese corpus to extend the ability of pointer network; modified pointer-based NMT model learn to point a word in the source sentence and copy it to the target sentence; post- process retranslating the raw translation by a phrase-based machine translation model or a wordlist. This framework can efficiently improve the problem of OOV and enhance the quality of translation of Uyghur-Chinese MT. Key Words: Uyghur-Chinese Machine Translation, deep neural network, deep learning, reordering model, out-of-vocabulary, Neural Machine Translation

文献类型学位论文
条目标识符http://ir.xjipc.cas.cn/handle/365002/4932
专题多语种信息技术研究室
作者单位中国科学院新疆理化技术研究所
推荐引用方式
GB/T 7714
孔金英. 深度神经网络在维汉机器翻译中的应用研究[D]. 北京. 中国科学院大学,2017.
条目包含的文件
文件名称/大小 文献类型 版本类型 开放类型 使用许可
深度神经网络在维汉机器翻译中的应用研究.(2587KB)学位论文 开放获取CC BY-NC-SA浏览 请求全文
个性服务
推荐该条目
保存到收藏夹
查看访问统计
导出为Endnote文件
谷歌学术
谷歌学术中相似的文章
[孔金英]的文章
百度学术
百度学术中相似的文章
[孔金英]的文章
必应学术
必应学术中相似的文章
[孔金英]的文章
相关权益政策
暂无数据
收藏/分享
文件名: 深度神经网络在维汉机器翻译中的应用研究.pdf
格式: Adobe PDF
此文件暂不支持浏览
所有评论 (0)
暂无评论
 

除非特别说明,本系统中所有内容都受版权保护,并保留所有权利。