机器翻译是指利用计算机自动的将一种自然语言翻译为另一种自然语言的技术。近年来，统计机器翻译与神经网络机器翻译已成为机器翻译研究领域的主流。统计机器翻译与神经网络机器翻译的基本思想是利用学习算法，从大规模双语句对中学习到翻译规则。但是，并不是所有的语言对之间都存在大规模的双语句对。为了解决机器翻译中的数据稀疏问题，研究人员提出了中轴语的概念，即在源语言与目标语中间引入中轴语，要求源语言与中轴语、中轴语与目标语之间存在大规模双语语料。本文针对基于中轴语的统计机器翻译，重点研究以下几个方面： 1.中轴语翻译的概率估计。目前，主流的中轴语翻译研究建立在基于短语的统计机器翻译基础上，这些研究通过直接将源语言-中轴语、中轴语-目标语的翻译概率相乘得到源语言-目标语的翻译概率。然而，由于源语言-中轴语、中轴语-目标语分布于两个不同的概率空间，并且可能存在着领域差异，这会导致翻译概率准确率低。为此，本文提出基于短语对共现次数的中轴语短语翻译模型，该方法首先抽取源语言-中轴语、中轴语-目标语短语，然后基于上述短语对推导出相应的源语言-目标语短语，最后在推导出的源语言-目标语短语对基础上计算源语言-目标语短语翻译概率。通过该模型，可以将原有的源语言-中轴语、中轴语-目标语翻译概率计算空间统一到源语言-目标语概率空间。在同一个概率空间下，可以得到更为准确的翻译概率。实验结果表明该方法在多个语言对上取得了比传统方法更优的结果。 2.中轴语翻译中的隐含翻译知识挖掘。作为统计机器翻译模型的载体，翻译短语表储存了翻译所需要的翻译知识，主要包含互译短语对、短语翻译概率等。对于目前基于短语的中轴语翻译方法，存在的一个缺陷是如果具有相同意思的源语言短语与目标语短语分别对应到不同的中轴语短语，那么对应的源语言与目标语短语对也会被丢失，这导致了翻译知识的丢失问题。为了解决这一问题，本文引入随机游走算法，在源语言-中轴语、中轴语-目标语短语表上构建翻译联通图，深度挖掘源语言-中轴语、中轴语-目标语短语表中的隐含翻译知识。实验结果表明随机游走算法可以有效地挖掘中轴语翻译中的隐含翻译知识，提高翻译效果。 3.中轴语翻译中的噪声过滤。模型中的噪声问题是统计机器翻译不可避免的问题。有多种因素可能导致噪声的产生，例如双语语料中的非对齐句对、词对齐过程中的错误等。在基于中轴语的统计机器翻译中，由于源语言-目标语短语表是由源语言-中轴语、中轴语-目标语短语表推导而来，这两个原始短语表中的噪声会被进一步放大到源语言-目标语短语表中。同时，由于中轴语短语本身可能带有多义性，这也会对源语言-目标语短语表带来新的噪声。为了解决中轴语翻译中的噪声问题，本文提出了基于最小贝叶斯风险的短语表过滤方法，这一方法在短语表中选择风险最小的短语进行过滤。实验结果表明这一方法可以在减小短语表规模的同时提升翻译性能。 4.中轴语翻译中的调序规则生成。对于传统统计机器翻译，研究人员通常使用词汇化调序模型来调整翻译结果中的短语顺序。词汇化调序模型将调序类型分为3种：单调、交换与不连续。生成一个词汇化调序模型需要依赖上下文信息，而在经典的基于中轴语的统计机器翻译中，上下文信息往往会被丢失。基于此，本文提出了基于扩展上下文信息的词汇化调序模型，通过将中轴语翻译中的上下文信息泛化为词性、词汇片段等信息，有效地解决了中轴语翻译中的调序问题。实验结果表明基于上下文信息的词汇化调序模型在多个语言方向上可以有效提高翻译效果。 关键词：统计机器翻译；中轴语；短语表；随机游走；调序；裁剪

 Machine translation refers to the use of computers to translate from one natural language into another natural language. Recent years, statistical machine translation and neural network machine translation have become the mainstream of machine translation research. The basic idea of statistical machine translation and neural network machine translation is to use learning algorithms to learn translation rules from large-scale bilingual corpus. However, large-scale bilingual corpus are not always available in some language pairs. To alleviate the data scarceness of machine translaiton, the pivot language approach is proposed as a “bridge” to connect the source and target language. The premise of the pivot approach is that a large number of source-pivot and pivot-target parallel data are available. In this paper, we focus on the following aspects of pivot-based statistical machine translation. 1.The probability estimation of pivot-based translation. The traditional pivot-based approach based on the phrase-based statistical machine translation, and proposes to build a source-target phrase table by merging the source-pivot and pivot-target phrase table. One of the key issues in this method is to estimate the translation probabilities for the generated source-target phrase pairs. Conventionally, the probabilities are estimated by multiplying the posterior probabilities of source-pivot and pivottarget phrase pairs. However, it has been shown that the generated probabilities are not accurate enough. One possible reason may lie in the nonuniformity of the probability space. To solve this problem, we propose a novel approach that utilizes the co-occurrence count of source-target phrase pairs to estimate phrase translation probabilities more precisely. Different from the triangulation method, which merges the source-pivot and pivot-target phrase pairs after training the translation model, we propose to merge the source-pivot and pivot-target phrase pairs immediately after the phrase extraction step, and estimate the cooccurrence count of the source-pivot-target phrase pairs. Finally, we compute the translation probabilities according to the estimated co-occurrence counts, using the standard training method in phrase-based SMT. Experimental results on Europarl data and web data show that our method leads to significant improvements over the baseline systems. 2.Hidden translation rules mining in pivot-based machine translation. One of the weaknesses of current pivot-based machine translation is that some corresponding source and target phrase pairs cannot be generated, because they are connected to different pivot phrases. To solve this problem, we apply a Markov random walk method to pivot-based SMT system to discover potential translations be-tween source and target language via the pivot language. Experimental results on Europarl corpus and web data show that our method leads to significant improvements over the baseline systems. 3.Noises and model pruning in pivot-based machine translation. The noise of the phrase table is a key problem in SMT. It is caused by many reasons, including: 1) the noises in parallel corpuses, 2) the defects of the learning algorithm, and so on. Because the pivot-based phrase table is generated by combining two standard SMT phrase table, the noises in a standard phrase table might be transferred and amplified in the pivot-based phrase table. Due to the ambiguities of the pivot language, source and target phrases with different meanings may be wrongly matched. Consequently, the derived source-target phrase table may contain incorrect phrase pairs. To alleviate this problem, we apply the minimum Bayes-risk method to prune the phrase table. The minimum Bayes-risk pruning method removes the phrase pairs with the lowest risk from the phrase table. Experimental results on Europarl data show that the proposed method can both reduce the size of phrase tables and improve the performance of translations. 4.The lexical reordering model in pivot-based machine translation. In phrase-based statistical machine translation, phrase reordering is a very important issue. Recent years, various phrase reordering methods are presented for SMT system to generate a fluent translation sentence. Among these reordering models, the lexical-ized reordering model is a commonly used method in current SMT systems. For each phrase pair, the lexicalized reordering model defines three types of orientations: directly follows a previous phrase (monotone), swapped with a previous phrase (swap), or not connected to the previous phrase (discontinuous). When applying the lexicalized reordering model into the triangulation method, a key problem is that the context information is missing in the phrase table. In this paper, we present a context-extended phrase reordering model for pivot-based statistical machine translation by extending the context information in source, pivot and target language. Experimental results show that our method leads to significant improvements over the baseline system. Keywords: statistical machine translation, pivot language, phrase table, random walk, reordering, model pruning