中国科学院自动化研究所机构知识库

近年来，机器翻译技术取得了长足的进步，译文质量不断提高，特别是基于端到端神经机器翻译方法(neural machine translation, NMT)的出现使得机器翻译的译文质量出现了变革式的发展，目前神经机器翻译方法已经成为机器翻译的新范式，同时各大公司依靠神经机器翻译方法改进了各自的机器翻译服务。然而，实践表明现有的神经机器翻译方法仍然存在一系列问题，而忠实度不高则是其中最为常见和严重的问题，更加具体地表现为错翻与漏翻现象。因此，研究增强译文忠实度的方法对提高机器翻译效果以及推动机器翻译的应用都具有重要的理论意义和应用价值。本文首先从分析现有神经机器翻译方法存在的问题出发，研究增强译文忠实度的神经机器翻译方法和实现技术。论文的主要工作和创新点归纳如下：

1、提出了一种融合词汇级翻译记忆的神经机器翻译方法

根据分析发现，某些类型的词语被神经机器翻译错翻的概率很高，本文称之为异常词。针对异常词，本文提出了融合词汇级翻译记忆的神经机器翻译方法以提升这类词语的翻译准确率。首先本文采取不同策略和标准来检测神经机器翻译系统的异常词。然后针对检测到的异常词，构建其词汇级翻译记忆存储每个异常词的候选译文及其翻译环境。最后通过一种动态的访问机制融合翻译记忆与神经机器翻译模型，以共同决定最终译文。实验表明，所提出的方法能够显著提高神经机器翻译系统的翻译效果，尤其能够显著减少神经翻译系统对于异常词的错翻率。

2、提出了一种融合短语级翻译记忆的神经机器翻译方法

词汇级翻译记忆缺乏上下文的约束，仍然存在一定的歧义性，然而统计机器翻译中的短语翻译规则直接编码了局部上下文信息，有助于消除歧义并且提升译文的忠实度。为此本文提出一种融合短语级翻译记忆的神经机器翻译方法。该方法的主要思想是在每个解码时刻利用短语翻译表构建推荐单词集，并提高推荐单词的预测概率。首先本文在短语翻译表中对源语言句子进行搜索并生成候选的目标短语集，随后将候选目标语言短语与已生成的部分译文进行匹配得到推荐单词集，然后计算每个推荐单词的推荐值，最后将推荐值与现有的神经机器翻译方法进行综合。实验表明，所提出的融合短语级翻译记忆的神经机器翻译方法能够充分利用短语翻译记忆来显著提高神经机器翻译的译文质量。

3、提出了一种基于翻译熵的神经机器翻译方法

除了错翻问题，漏翻也是导致神经机器翻译系统译文忠实度不高的关键因素。针对漏翻现象，本文提出基于翻译熵的神经机器翻译方法。首先本文通过实验分析发现源端单词的漏翻率与其翻译熵密切相关：一个单词的翻译熵越高，其漏翻率也越高。为了缓解高熵词的漏翻问题，本文进而提出一种基于翻译熵的神经机器翻译方法以减少高熵词的漏翻现象。所提出的方法是一种从粗粒度到细粒度的框架：在粗粒度阶段，构造一种伪目标语言来泛化高熵词语，以减少这类词语由于翻译不确定性而导致的漏翻现象；在细粒度阶段，利用构造的伪目标语言来提高现有神经机器翻译系统的性能。实验表明，所提出的方法能够有效地减少高熵词的漏翻现象。

综上所述，本文针对神经机器翻译存在的忠实度不高问题展开了深入研究，主要关注神经机器翻译中存在的错翻和漏翻问题，并分别提出了减少错翻和漏翻的方法。最终实验证实本文所提出的方法能够有效减少神经机器翻译的错翻和漏翻现象，相关成果有力地推动了神经机器翻译的研究与应用。

In recent years, the research on machine translation has made considerable progress and the performance of machine translation has been improved a lot. Especially, the end-to-end Neural Machine Translation (NMT) method has led to a revolutionary progress in the translation quality. At present, NMT method has become a new paradigm of machine translation. And many companies have offered better translation service with the help of NMT method. While statistics show that the current NMT methods still have several drawbacks, and the most common and serious one is the low-faithfulness problem, which is manifested as the mis-translation and under-translation. Therefore, studying the methods to improve the faithfulness of NMT has important theoretical significance and application value. Therefore, this paper first investigates the drawback of the current NMT models, then proposed several methods to improve the faithfulness of NMT models. The main contributions of this paper are summarized as follows:

1. Incorporating Lexical-level Translation Memory into Neural Machine Translation

Statistics show that the mis-translation ratio of some words is much higher, and we refer these words as troublesome words. To address the problem of troublesome words, this paper proposes a method to incorporate lexical-level translation memory into NMT. In this paper, we first investigate three different strategies to define and detect the troublesome words. We then constructed a contextual memory to memorize which target words should be produced in what situations. Finally, to correctly translate the troublesome words, we design a hybrid model to dynamically access the contextual memory and combine the memory with NMT model. The extensive experiments demonstrate that our method can significantly outperform the baseline models in translation quality, especially in handling troublesome words.

2. Incorporating Phrasal-level Translation Memory into Neural Machine Translation

Due to the lack of the context constraints, lexical-level translation memory may lead to some ambiguities. However, the phrasal-level translation table in statistical machine translation (SMT) can directly encode local context information, which may help to eliminate ambiguities and enhance the faithfulness of the translation. Therefore, in the paper, we propose a neural machine translation method with phrasal-level translation memory. The main idea is to add bonus to words worthy of recommendation, so that NMT can make correct predictions. Specifically, we first derive a prefix tree to accommodate all the candidate target phrases by searching the phrase translation table according to the source sentence.
Then, we construct a recommendation word set by matching between candidate target phrases and previously translated target words by NMT. After
that, we determine the specific bonus value for each recommendable word. Finally,
we integrate this bonus value into NMT. The extensive experiments demonstrate that the proposed methods obtain remarkable improvements over the attention based NMT.

3. Proposing Translation-Entropy based Neural Machine Translation

Besides the mis-translation, the low-faithfulness problem of the current NMT is also reflected in the under-translation. In this paper, we propose a translation-entropy based neural machine translation method to address the this problem. Through analysis, we observe that a source word with a large translation entropy is more inclined to be dropped. Based on this, we propose a coarse-to-fine framework. In coarse-grained phase, we introduce a simple strategy to reduce the entropy of high entropy words through constructing the pseudo target sentences. In fine-grained phase, we propose three methods to improve the NMT with the help of the pseudo target sentences. Experimental results show that our method can significantly improve the translation quality and substantially reduce the under-translation cases of high-entropy words.

In summary, this paper aims to address the low-faithfulness problem in NMT model, and mainly focuses on the mis-translation and under-translation problem. We proposed several methods to address the mis-translation and under-translation problem. Finally, the experiments show that the proposed method can effectively reduce the cases of mis-translation and under-translation in NMT. The relevant methods strongly promote the research and application of NMT models.