{
    "paper_id": "I17-1039",
    "header": {
        "generated_with": "S2ORC 1.0.0",
        "date_generated": "2023-01-19T07:38:57.648452Z"
    },
    "title": "Towards Neural Machine Translation with Partially Aligned Corpora",
    "authors": [
        {
            "first": "Yining",
            "middle": [],
            "last": "Wang",
            "suffix": "",
            "affiliation": {
                "laboratory": "National Laboratory of Pattern Recognition",
                "institution": "CASIA",
                "location": {
                    "settlement": "Beijing",
                    "country": "China"
                }
            },
            "email": "yining.wang@nlpr.ia.ac.cn"
        },
        {
            "first": "Yang",
            "middle": [],
            "last": "Zhao",
            "suffix": "",
            "affiliation": {
                "laboratory": "National Laboratory of Pattern Recognition",
                "institution": "CASIA",
                "location": {
                    "settlement": "Beijing",
                    "country": "China"
                }
            },
            "email": "yang.zhao@nlpr.ia.ac.cn"
        },
        {
            "first": "Jiajun",
            "middle": [],
            "last": "Zhang",
            "suffix": "",
            "affiliation": {
                "laboratory": "National Laboratory of Pattern Recognition",
                "institution": "CASIA",
                "location": {
                    "settlement": "Beijing",
                    "country": "China"
                }
            },
            "email": "jjzhang@nlpr.ia.ac.cn"
        },
        {
            "first": "Chengqing",
            "middle": [],
            "last": "Zong",
            "suffix": "",
            "affiliation": {
                "laboratory": "National Laboratory of Pattern Recognition",
                "institution": "CASIA",
                "location": {
                    "settlement": "Beijing",
                    "country": "China"
                }
            },
            "email": "cqzong@nlpr.ia.ac.cn"
        },
        {
            "first": "Zhengshan",
            "middle": [],
            "last": "Xue",
            "suffix": "",
            "affiliation": {
                "laboratory": "",
                "institution": "Toshiba (China) Co.,Ltd",
                "location": {}
            },
            "email": "xuezhengshan2@toshiba.com.cn"
        }
    ],
    "year": "",
    "venue": null,
    "identifiers": {},
    "abstract": "While neural machine translation (NMT) has become the new paradigm, the parameter optimization requires large-scale parallel data which is scarce in many domains and language pairs. In this paper, we address a new translation scenario in which there only exists monolingual corpora and phrase pairs. We propose a new method towards translation with partially aligned sentence pairs which are derived from the phrase pairs and monolingual corpora. To make full use of the partially aligned corpora, we adapt the conventional NMT training method in two aspects. On one hand, different generation strategies are designed for aligned and unaligned target words. On the other hand, a different objective function is designed to model the partially aligned parts. The experiments demonstrate that our method can achieve a relatively good result in such a translation scenario, and tiny bitexts can boost translation quality to a large extent.",
    "pdf_parse": {
        "paper_id": "I17-1039",
        "_pdf_hash": "",
        "abstract": [
            {
                "text": "While neural machine translation (NMT) has become the new paradigm, the parameter optimization requires large-scale parallel data which is scarce in many domains and language pairs. In this paper, we address a new translation scenario in which there only exists monolingual corpora and phrase pairs. We propose a new method towards translation with partially aligned sentence pairs which are derived from the phrase pairs and monolingual corpora. To make full use of the partially aligned corpora, we adapt the conventional NMT training method in two aspects. On one hand, different generation strategies are designed for aligned and unaligned target words. On the other hand, a different objective function is designed to model the partially aligned parts. The experiments demonstrate that our method can achieve a relatively good result in such a translation scenario, and tiny bitexts can boost translation quality to a large extent.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Abstract",
                "sec_num": null
            }
        ],
        "body_text": [
            {
                "text": "Neural machine translation (NMT) proposed by Kalchbrenner et al.(2013) , Sutskever et al.(2014) and Cho et al.(2014) has achieved significant progress in recent years. Different from traditional statistical machine translation(SMT) (Koehn et al., 2003; Chiang, 2005; Liu et al., 2006; Zhai et al., 2012) which contains multiple separately tuned components, NMT builds an end-to-end framework to model the whole translation process. For several language pairs, NMT is reaching significantly better translation performance than SMT (Luong et al., 2015b; .",
                "cite_spans": [
                    {
                        "start": 45,
                        "end": 70,
                        "text": "Kalchbrenner et al.(2013)",
                        "ref_id": "BIBREF11"
                    },
                    {
                        "start": 73,
                        "end": 95,
                        "text": "Sutskever et al.(2014)",
                        "ref_id": "BIBREF27"
                    },
                    {
                        "start": 100,
                        "end": 116,
                        "text": "Cho et al.(2014)",
                        "ref_id": "BIBREF5"
                    },
                    {
                        "start": 232,
                        "end": 252,
                        "text": "(Koehn et al., 2003;",
                        "ref_id": "BIBREF13"
                    },
                    {
                        "start": 253,
                        "end": 266,
                        "text": "Chiang, 2005;",
                        "ref_id": "BIBREF4"
                    },
                    {
                        "start": 267,
                        "end": 284,
                        "text": "Liu et al., 2006;",
                        "ref_id": "BIBREF16"
                    },
                    {
                        "start": 285,
                        "end": 303,
                        "text": "Zhai et al., 2012)",
                        "ref_id": "BIBREF31"
                    },
                    {
                        "start": 530,
                        "end": 551,
                        "text": "(Luong et al., 2015b;",
                        "ref_id": "BIBREF19"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "In general, in order to obtain an NMT model \u5916\u4ea4\u90e8\u53d1\u8a00\u4eba\u5218\u5efa\u8d85\u4eca\u5929\u5728\u4f8b\u884c\u7684\u673a\u5236\u62db\u5f85\u4f1a\u4e0a\u8bf4\uff0c \u7f8e\u56fd\u53f8\u6cd5\u90e8\u957f\u6765\u5317\u4eac\u8fdb\u884c\u4e86\u8bbf\u95ee\u3002 Speaking at a regular press briefing Wednesday, Turkish foreign ministry deputy spokesman said that turkey hoped the peace talks would continue as planned.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Introduction",
                "sec_num": "1"
            },
            {
                "text": "Partial Aligned Part 1 Figure 1 : An example of our partially aligned training data, in which the source sentence and target sentence are not parallel but they include two parallel parts (highlight in blue and red respectively).",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 23,
                        "end": 31,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Partial Aligned Part 2",
                "sec_num": null
            },
            {
                "text": "of great translation quality, we usually need largescale parallel data. Unfortunately, the large-scale parallel data is always insufficient in many domains and language pairs. Without sufficient parallel sentence pairs, NMT tends to learn poor estimates on low-count events. Actually, there have been some effective methods to deal with the situation of translating language pairs with limited resource under different scenarios (Johnson et al., 2016; Sennrich et al., 2016a; . In this paper, we address a new translation scenario in which we do not have any parallel sentences but have massive monolingual corpora and phrase pairs. The previous methods are hard to be used to learn an NMT model under this situation. In this paper, we propose a novel method to learn an NMT model using only monolingual data and phrase pairs.",
                "cite_spans": [
                    {
                        "start": 429,
                        "end": 451,
                        "text": "(Johnson et al., 2016;",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 452,
                        "end": 475,
                        "text": "Sennrich et al., 2016a;",
                        "ref_id": "BIBREF24"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Partial Aligned Part 2",
                "sec_num": null
            },
            {
                "text": "Our main idea is that although there does not exist the parallel sentences, we can derive the sentence pairs which are non-parallel but contain the parallel parts (in this paper, we call these sentences as partially aligned sentences) with the monolingual data and phrase pairs. Then we can utilize these partially aligned sentences to train an NMT model. Figure 1 shows an example of our data. Source sentence and target sentence are not fully aligned but contain two translation fragments: (\"\u5916\u4ea4\u90e8\u53d1\u8a00\u4eba\", \"foreign ministry deputy\") and (\"\u5728 \u4f8b \u884c \u7684 \u8bb0 \u8005 \u62db \u5f85 \u4f1a \u4e0a \u8bf4\", \"speaking at a regular press\"). Intuitively, these kinds of sentence pairs are useful in building an NMT model.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 356,
                        "end": 364,
                        "text": "Figure 1",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Partial Aligned Part 2",
                "sec_num": null
            },
            {
                "text": "To use these partially aligned sentences, the training method should be different from the original methods which are designed for parallel corpora. In this work, we adapt the conventional NMT training method mainly from two perspectives. On one hand, different generation strategies are designed for aligned and unaligned target words. For aligned words, our method guides the translation process based on both the context of source side and previously predicted words. When generating the unaligned target words , our model only depends on the words previously generated without considering the context of source side. On the other hand, we redesign the objective function so as to emphasize the partially aligned parts in addition to maximizing the log-likelihood of the target sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Partial Aligned Part 2",
                "sec_num": null
            },
            {
                "text": "The contributions of our paper are twofold: 1) Our approach addresses a new translation scenario, where there only exists monolingual data and phrase pairs. We propose a method to train an NMT model under this scenario. The method is simple and easy to implement, which can be used in arbitrary attention-based NMT framework.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Partial Aligned Part 2",
                "sec_num": null
            },
            {
                "text": "2) Empirical experiments on the Chinese-English translation tasks under this scenario show that our method can achieve a relatively good result. Moreover, if we only add a tiny parallel corpus, the method can obtain significant improvements in terms of translation quality.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Partial Aligned Part 2",
                "sec_num": null
            },
            {
                "text": "Our approach can be easily applied to any endto-end attention-based NMT framework. In this work, we follow the neural machine translation architecture by Bahdanau et al. (2015) , which we will summarize in this section. Given the source sentence X = {x 1 , x 2 , ..., x T x } and the target sentence Y = {y 1 , y 2 , ..., y T y }. The goal of machine translation is to transform source sentence into the target sentence. The end-to-end NMT framework consists of two recurrent neural networks, which are respectively called encoder and decoder. First, the encoder network encodes the X into context vectors C. Then, the decoder network generates the target translation sentences one word each time based on the context vectors C and target words previously generated. More specifically, that is p(y i |y <i , C).",
                "cite_spans": [
                    {
                        "start": 154,
                        "end": 176,
                        "text": "Bahdanau et al. (2015)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "In encoding stage, it transforms X into a sequence of vectors",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "h enc = h k 1 , h k 2 , h k 3 , .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": ".., h k T using m stacked LSTM (Hochreiter and Schmidhuber, 1997) layers. Finally, the encoder chooses the hidden states of the top encoder layer as",
                "cite_spans": [
                    {
                        "start": 31,
                        "end": 65,
                        "text": "(Hochreiter and Schmidhuber, 1997)",
                        "ref_id": "BIBREF9"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "h top = {h m 1 , h m 2 , h m 3 , .",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": ".., h m T } that we will use in attention mechanism to calculate context vector later.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "In decoding stage, it generates one target word at a time from conditional probability p(y i |y <i , C; \u03b8) also via m stacked LSTM layers parameterized by \u03b8. Supposing we have obtained the context vector, the conditional probability p(y i |y <i , C; \u03b8) is calculated as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "p(y i |y <i , C; \u03b8) = p(y i |y <i , c i ) = sof tmax(g(W y i , z m i , c i ))",
                        "eq_num": "(1)"
                    }
                ],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "Where W y i is embedding of the target word, z m i is current hidden states of top layer in decoder network. Note that the first hidden states of decoder z k 0 are set to the last hidden states of encoder as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "z k 0 = h k T",
                        "eq_num": "(2)"
                    }
                ],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "c i can be computed as a weighted sum of the source-side h s as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "c i = Tx j=1 a ij h m i",
                        "eq_num": "(3)"
                    }
                ],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "Where a ij is alignment probability, which can be calculated in multiple ways (Luong et al., 2015a) . In our method, we use a simple singlelayer feed forward network. This alignment probability measures how relevant i-th context vector of source sentence is in deciding the current symbol in translation. The probability will be further normalized:",
                "cite_spans": [
                    {
                        "start": 78,
                        "end": 99,
                        "text": "(Luong et al., 2015a)",
                        "ref_id": "BIBREF18"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "a ij = exp(e ij ) T x k=1 exp(e ik ))",
                        "eq_num": "(4)"
                    }
                ],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "A detailed introduction of the encoder-decoder framework is described in Bahdanau et al. (2015) . In order to train NMT system, we use parallel data to optimize the network parameters by maximizing the conditional log-likelihood:",
                "cite_spans": [
                    {
                        "start": 73,
                        "end": 95,
                        "text": "Bahdanau et al. (2015)",
                        "ref_id": "BIBREF0"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "L(\u03b8, D) = 1 N N n=1 Ty i log(p(y (n) i |p(y (n) <i , X (n) , \u03b8))",
                        "eq_num": "(5)"
                    }
                ],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "3 NMT with Partially Aligned Data",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "In \u00a72 we gave a brief description of the attentionbased NMT models whose network parameters are trained using parallel sentence pairs. However, in the translation scenario where there only exists monolingual corpora and phrase pairs, the conventional NMT framework is hard to be used in training a model. In this section, we first explain how we actually obtain the partially aligned corpora with aligned positions using phrase pairs and monolingual corpora, then introduce our method to train the NMT models using the partially aligned sentences according to the particular properties of the corpora.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Review of Neural Machine Translation",
                "sec_num": "2"
            },
            {
                "text": "Assuming there exists abundant phrase pairs and monolingual sentences in source and target languages, we define our approach to extract partially aligned sentence pairs for training. Given a phrase pair (ph s, ph t), ph s may appear in a source-side monolingual sentence X, and ph t may appear in a target-side monolingual sentence Y. Then, X and Y are non-parallel but contain the parallel part. We call these kinds of data the partially aligned sentences. In this work, we collect the partially aligned sentences by searching the phrase pairs in both of the source and target monolingual data simultaneously. In order to reduce the time of the searching process, the monolingual training corpora are first split into many parts. Then, we retrieve the source phrase in each part to restrict the source range of partially aligned corpus. With the retrieved results, we can search the final results of the partially aligned sentences pairs easily. In this way we can construct our corpora, in which only one or more phrases are aligned in every sentence pairs. We denote a partially aligned sentence (",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constructing partially Aligned Corpora",
                "sec_num": "3.1"
            },
            {
                "text": "X = x 1 , x 2 , ...x T x ,Y = y 1 , y 2 , ..., y T y ),",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constructing partially Aligned Corpora",
                "sec_num": "3.1"
            },
            {
                "text": "in which a set of the phrase pairs aligned with each other. We call these pairs partially aligned part:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constructing partially Aligned Corpora",
                "sec_num": "3.1"
            },
            {
                "text": "P (k) x = x k1 , ..., x kp P (k) y = y k1 , ..., x kq (6) P (k)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constructing partially Aligned Corpora",
                "sec_num": "3.1"
            },
            {
                "text": "x and P (k)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constructing partially Aligned Corpora",
                "sec_num": "3.1"
            },
            {
                "text": "y are the phrases in the source and target sentences respectively, and they are translation equivalents.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Constructing partially Aligned Corpora",
                "sec_num": "3.1"
            },
            {
                "text": "In \u00a73.1, we acquired the partially aligned corpora with the phrase pairs and monolingual sentences. Now, we need to use them to train the NMT model. As the traditional NMT model is designed for the parallel sentences, it is not suitable for partially aligned sentences. Thus we redesign the traditional NMT model as follows. Figure 2 shows the basic framework of our training method. Our model has 4 different parts from conventional NMT model, including initial states, generation process, objective functions and vocabulary size.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 325,
                        "end": 333,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Model Descriptions",
                "sec_num": "3.2"
            },
            {
                "text": "The first difference is the initial hidden states of decoder. In the conventional NMT model, the initial hidden states of decoder z k 0 is set to the last hidden state in encoder h k T x , including initial states, generation process, objective functions and vocabulary size. shown in Eq. (2). For parallel sentences, this setting is reasonable, while for partially aligned sentence, this initial method is inappropriate. The reason is that the hidden state of last word in source sentence is irrelevant to the target sentence, considering the fact that under our scenario, the target side sentences may entirely uncorrelated to the source sentences. Thus, in our model, z k 0 is set to a zero vector as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Initial States",
                "sec_num": "3.2.1"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "Z k 0 = 0",
                        "eq_num": "(7)"
                    }
                ],
                "section": "Initial States",
                "sec_num": "3.2.1"
            },
            {
                "text": "In Figure 2 , the hidden state of \"<start>\" symbol is set to zero vector when y 1 does not belong to parallel part of the sentence pair.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 3,
                        "end": 11,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Initial States",
                "sec_num": "3.2.1"
            },
            {
                "text": "The second difference is the generation process. In the conventional NMT system, the model generates each target word based on the context vector c i and previously predicted words y <i as shown in Eq. (1). This generation strategy is unsuitable for the partially aligned corpora, since there exists many unaligned target words. Intuitively, when the model generates the non-parallel parts, it is unnecessary to take the context vector c i into consideration. As Figure 2 illustrated, (x 2 ,x 3 ) and (y 2 ,y 3 ) are parallel parts in this partially aligned sentence pair, and we use context vector which is generated by attention mechanism only when the decoder outputs y 2 and y 3 . Therefore, the decoder in our model can be described as follows:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 463,
                        "end": 471,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Generation Process",
                "sec_num": "3.2.2"
            },
            {
                "text": "c i = T x j a ij h i if y i \u2208 P (k) y 0 if y i / \u2208 P (k) y (8)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generation Process",
                "sec_num": "3.2.2"
            },
            {
                "text": "where a ij is calculated as Eq. (2), P",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generation Process",
                "sec_num": "3.2.2"
            },
            {
                "text": "(k) y",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generation Process",
                "sec_num": "3.2.2"
            },
            {
                "text": "is the target partial part, as shown in Eq. (5). In Eq. (8), our model generates the aligned target words based on the context vector c i and previously predicted words y <i . When generating the unaligned target words, the model sets the context vector c i to zero, indicating that the model generates these words only based on the LSTM-based RNN language model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Generation Process",
                "sec_num": "3.2.2"
            },
            {
                "text": "Third, we redesign the objective function. Given the parallel data, the objective function is to maximize the log-likelihood of each target word as shown in Eq. (5). For the partially aligned sentence, besides the source and target sentence, we know the phrase alignment information. Hence, apart from maximizing the log-likelihood of each target word, we also hope to make the source and target words in partially aligned part align to each other. As shown in Figure 2 , when predicting the words y 2 and y 3 , we want to attend more information of corresponding words x 2 and x 3 . Thus, we inject an auxiliary object function to achieve it. More specifically, our objective function is de-signed as follows:",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 461,
                        "end": 469,
                        "text": "Figure 2",
                        "ref_id": "FIGREF0"
                    }
                ],
                "eq_spans": [],
                "section": "Objective Function",
                "sec_num": "3.2.3"
            },
            {
                "text": "L P (\u03b8, D) = 1 N N n=1 { Ty i log(p(y (n) i |p(y (n) <i , X (n) ; \u03b8)) + \u03bb \u00d7 (a (n) , a (n) ); \u03b8)} (9)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Objective Function",
                "sec_num": "3.2.3"
            },
            {
                "text": "Where a (n) is defined in Eq. 4, is a loss function that encourages the agreement between a (n) and a (n) . a (n) is the supervised attention determined by alignment relationship between P X and P Y , and can be calculated as follows:",
                "cite_spans": [
                    {
                        "start": 102,
                        "end": 105,
                        "text": "(n)",
                        "ref_id": null
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Objective Function",
                "sec_num": "3.2.3"
            },
            {
                "text": "a (n) i,j = 1 if x j \u2208 P X and y i \u2208 P Y 0 others (10)",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Objective Function",
                "sec_num": "3.2.3"
            },
            {
                "text": "\u03bb > 0 is a hyper-parameter that balances the preference between likelihood and agreement. In this paper, it is set to 0.3. As shown in Eq. 8, our objective function does not only consider to maximize the loglikelihood of the target sentence, but also encourages the alignment a ij produced by NMT to have a larger agreement with the prior alignment information. This objective function is similar to that used by the supervised attention method (Mi et al., 2016a; . Inspired by , the agreement between a (n) and a (n) can be defined in different ways:",
                "cite_spans": [
                    {
                        "start": 445,
                        "end": 463,
                        "text": "(Mi et al., 2016a;",
                        "ref_id": "BIBREF21"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Objective Function",
                "sec_num": "3.2.3"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "\u2022 Multiplication (MUL) (a (n) , a (n) i,j ; \u03b8) = \u2212 T y i=0 T x j=0 a(\u03b8) (n) i,j \u00d7 a (n) i,j (11) where a (n) i,j is computed by Eq. (10) \u2022 Mean Squared Error (MSE) (a (n) , a (n) i,j ; \u03b8) = \u2212 T y i=0 T x j=0 1 2 (a(\u03b8) (n) i,j \u2212 a (n) i,j ) 2",
                        "eq_num": "(12)"
                    }
                ],
                "section": "Objective Function",
                "sec_num": "3.2.3"
            },
            {
                "text": "The last difference is the vocabulary size during decoding. To make use of phrase pairs as much as possible, we extract a number of special phrase pairs whose source and target are both one word.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Limited Vocabulary",
                "sec_num": "3.2.4"
            },
            {
                "text": "In decoding phase, as what Mi et al. (2016b) have done, we can use these special phrase pairs to reduce the vocabulary size when computing the final score distribution. In this way, we can not only acquire more accurate translation of each word, but also accelerate the decoding speed. The vocabulary size can be reduced as follows:",
                "cite_spans": [
                    {
                        "start": 27,
                        "end": 44,
                        "text": "Mi et al. (2016b)",
                        "ref_id": "BIBREF22"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Limited Vocabulary",
                "sec_num": "3.2.4"
            },
            {
                "text": "EQUATION",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [
                    {
                        "start": 0,
                        "end": 8,
                        "text": "EQUATION",
                        "ref_id": "EQREF",
                        "raw_str": "V = V 1 \u222a V 2",
                        "eq_num": "(13)"
                    }
                ],
                "section": "Limited Vocabulary",
                "sec_num": "3.2.4"
            },
            {
                "text": "Where V 1 contains the most frequently target words and V 2 is a target words set. This set V 2 is made up of all the target words of the special phrase pairs whose corresponding source words belong to the source sentence.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Limited Vocabulary",
                "sec_num": "3.2.4"
            },
            {
                "text": "In this section, we perform the experiment on Chinese-English translation tasks to test our method.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Experiment",
                "sec_num": "4"
            },
            {
                "text": "We evaluate our approach on large-scale monolingual data set from LDC corpus, which includes 13M Chinese sentences and 10M English sentences. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Dataset",
                "sec_num": "4.1"
            },
            {
                "text": "Considering the fact that the amount of manually annotated phrase pairs is not enough, in order to imitate the environment of experiment, we extract phrase pairs from parallel corpora automatically to make up for the shortage of quantity. To do this, we use Moses (Koehn et al., 2007) in its training step to learn a phrase table from LDC corpus, which includes 0.63M sentence pairs. In order to simulate the experiment as far as possible, we adopt three strategies to filter low quality phrase pairs: 1) the phrases containing the punctuation should be filtered out. (The special phrase parirs introduced in \u00a73.2.4 should be retain) 2) the length of source phrase and target phrase should be greater than 3. 3) only the phrase pairs whose translation probability exceed 0.5 should be retain. In this way, we can get 3M phrase pairs in our experiment. According to our analysis, the average length of phrases are 4.15 and 4.70 on source and target side respectively. When we search the phrase pairs in monolingual sentences, an obstacle is that one phrase pair will get different source sentences with same target sentence or same source sentence with different target sentences. Therefore, for one phrase pair, we have to restrict the number n of both source sentences and target sentences. To balance the search speed of the phrase pairs in monolingual corpora and the amount of partially aligned sentences, we set the hyper-parameter n to 7. We can search for 5M partially aligned sentences in our experiment. We also calculate the average length ratio of aligned phrases against the whole sentence, which is only 21% and 23% respectively on source and target side.",
                "cite_spans": [
                    {
                        "start": 264,
                        "end": 284,
                        "text": "(Koehn et al., 2007)",
                        "ref_id": "BIBREF12"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data Preparing and Preprocessing",
                "sec_num": "4.2"
            },
            {
                "text": "To ensure the quality of the partially aligned corpora, we also set the number of phrases that aligned to each other in one sentence pair must be greater than a threshold. Here, the threshold is set to 2. That is to say the partially aligned sentence pair should contain at least two aligned phrase pairs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Data Preparing and Preprocessing",
                "sec_num": "4.2"
            },
            {
                "text": "We build our described method based on the Zoph RNN toolkit 1 written in C++/CUDA. Both encoder and decoder consist of two stacked LSTM layers. We set minibatch size to 128. The word embedding dimension of both source and target sides is 1000, and the dimensions of hidden layers unit is set to 1000. In our baseline model, we limit the vocabulary of both source and target languages to 30K most frequent words, and other words are replaced by a special symbol \"UNK\". We run our model on the training corpus 20 iterations in total with stochastic gradient decent algorithm. We set learning rate to 0.1 at the beginning and halve the threshold while the perplexity increases on the development set. Dropout is applied to our model, and the rate is set to 0. ",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Details",
                "sec_num": "4.3"
            },
            {
                "text": "We conduct our experiment on the dataset mentioned above, and we list the training methods used as follows:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Methods",
                "sec_num": "4.4"
            },
            {
                "text": "1) Phrase NMT Model: As mentioned above, the only parallel resource is phrase pairs. We use attention-based NMT system to train only on the 3M phrase pairs to get our baseline result.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Methods",
                "sec_num": "4.4"
            },
            {
                "text": "2) Partially Aligned Model(MUL):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Methods",
                "sec_num": "4.4"
            },
            {
                "text": "We train our NMT model using the objective function of multiplication method on the partially aligned sentences.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Methods",
                "sec_num": "4.4"
            },
            {
                "text": "3) Partially Aligned Model(MSE):",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Methods",
                "sec_num": "4.4"
            },
            {
                "text": "We train our NMT model using the objective function of Mean Squared Error(MSE) method.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Training Methods",
                "sec_num": "4.4"
            },
            {
                "text": "Partially Aligned Model(MSE) + Limited-Vocab: It is similar to Partially Aligned Model(MSE) and the only difference is that we restrict the final score distribution on a limited target vocabulary, which is described in \u00a73.2.4.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "4)",
                "sec_num": null
            },
            {
                "text": "Phrase NMT Model + LimitedVocab: It is the method that LimitedVocab is used in Phrase NMT model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "5)",
                "sec_num": null
            },
            {
                "text": "We present the translation results in BLEU scores of different systems in Table 2 . Our first concern is whether the proposed model can actually improve the translation quality. As Table 2 shows, we find that our partially aligned model (both MUL supervised method and MSE supervised method) is superior to the Phrase NMT Model, which indicates that our Partially Aligned method is effective in improving the translation quality.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 74,
                        "end": 81,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    },
                    {
                        "start": 181,
                        "end": 188,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Phrase NMT Model vs. Partially Aligned Method",
                "sec_num": "5.1"
            },
            {
                "text": "\u5e03\u4ec0\u8bf4\uff1a\"\u6b64\u9879\u8ba1\u5212\u5c06\u5bf9\u52b3\u52a8\u5927\u4f17\u63d0\u4f9b \u51cf\u7a0e\u4f18\u60e0\u3002\"",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phrase NMT Model vs. Partially Aligned Method",
                "sec_num": "5.1"
            },
            {
                "text": "The bush says the <UNK> tax concessions be made a possible member Bush said, the scheme will provide tax concessions to the community.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phrase NMT Model vs. Partially Aligned Method",
                "sec_num": "5.1"
            },
            {
                "text": "Source sentence:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phrase NMT Model vs. Partially Aligned Method",
                "sec_num": "5.1"
            },
            {
                "text": "Phrase NMT model:",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phrase NMT Model vs. Partially Aligned Method",
                "sec_num": "5.1"
            },
            {
                "text": "Our model: Figure 3 : A translation example comparing between phrase NMT model and our partially aligned method. Figure 3 lists a comparison example of Phrase NMT model and our model. Obviously, our model can achieve the correct translation while the Phrase NMT model generates the unfaithful result. It demonstrates that our model is actually having the ability to learn translation adequacy from aligned parts and fluency from both aligned and unaligned parts.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 11,
                        "end": 19,
                        "text": "Figure 3",
                        "ref_id": null
                    },
                    {
                        "start": 113,
                        "end": 121,
                        "text": "Figure 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Phrase NMT Model vs. Partially Aligned Method",
                "sec_num": "5.1"
            },
            {
                "text": "In \u00a73.2, we described two objective functions in our model. We focus on the difference of these two approaches. As a comparison, MSE method (Row 3) outperforms the MUL method (Row 2) with an average improvement of 0.8 BLEU points, indicating that MSE method is more effective as an object function for our partially aligned model. In the following experiment, we adopt the MSE method as the new objective function used in our model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Phrase NMT Model vs. Partially Aligned Method",
                "sec_num": "5.1"
            },
            {
                "text": "In Table 2 , an interesting result is that using a reduced vocabulary can significantly improve the performance (+1.36 BLEU points), but it can only achieve 0.12 BLEU points improvement for Phrase NMT model. According to Mi et al. (2016b) , this approach is useful in conventional NMT model. Our result is in agreement with their findings, and the improvement is more prominent in our partially aligned model. Under our scenario, compared to the parallel corpora, fewer parallel parts appear in sentence pairs. The faithfulness of our translation result is relatively poor while the fluency is relatively good. With the limited relevant vocabulary, the faithfulness of the translation results is much improved. ",
                "cite_spans": [
                    {
                        "start": 221,
                        "end": 238,
                        "text": "Mi et al. (2016b)",
                        "ref_id": "BIBREF22"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 3,
                        "end": 10,
                        "text": "Table 2",
                        "ref_id": "TABREF2"
                    }
                ],
                "eq_spans": [],
                "section": "Effect of Limited Vocabulary",
                "sec_num": "5.2"
            },
            {
                "text": "The performance of our partially aligned model with different lengths is another problem we care about. We randomly select 1000 sentences from translation results of test set (MT03-08), which are trained by the Partially Aligned Model(MSE)+ReduceDict method. We classify them into five categories according to the length. Figure 4 shows the results of the experiment.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 322,
                        "end": 330,
                        "text": "Figure 4",
                        "ref_id": "FIGREF1"
                    }
                ],
                "eq_spans": [],
                "section": "Result of Different Sentence Lengths",
                "sec_num": "5.3"
            },
            {
                "text": "We find that the sentences with shorter length (lower than 40) yield better results than the long sentences. When the length of sentences exceeds 80, the quality of translation is rather poor. The reason is that we mainly use phrase pairs in our method, and the length of phrase is relative short compared to whole sentence. So our model is more suitable for the translation of short sentences. When we translate long sentences, the parameters in our model are not adjusted for tuning, and our approach can not produce translation of high quality.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Result of Different Sentence Lengths",
                "sec_num": "5.3"
            },
            {
                "text": "We concern that when we have a tiny parallel corpus, whether the small scale parallel corpus can boost the translation performance of the partially aligned method. Here, we fine-tune the partially aligned translation model on these corpora. The details of these corpora are introduced in Table 3 .",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 288,
                        "end": 295,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Effect of Adding Small Parallel Corpus",
                "sec_num": "5.4"
            },
            {
                "text": "Chinese English parallel #Sent. 0.1M #Word 3.00M",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Corpus",
                "sec_num": null
            },
            {
                "text": "3.86M Vocab 0.07M 0.04M Table 3 : The statistics of small-scale parallel datasets.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 24,
                        "end": 31,
                        "text": "Table 3",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Corpus",
                "sec_num": null
            },
            {
                "text": "The result is presented in Table 4 . We can observe that the translation result tends to be poor by only using a small-scale parallel corpora. It indicates that conventional NMT system cannot learn a good model on the small-scale datasets. However, when fine-tuning our partially aligned model with this small parallel corpus, we can get a surprising improvement. The results suggest that when under a scenario in which we only have monolingual corpora and phrase pairs, even a few bitexts can boost translation quality to a large extent.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 27,
                        "end": 34,
                        "text": "Table 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Corpus",
                "sec_num": null
            },
            {
                "text": "We investigate the effect of the different corpora size on the final translation results. According to Table 4 , when the number of parallel sentences is quite small (lower than 60K), we can acquire a measurable improvement (more than 10 BLEU) compared to the conventional NMT system result. Especially, when the size of sentence pairs is 40K and 60K, our method obtains the enormous improvement over the NMT model by +13.82 BLEU points and +13.21 BLEU points respectively. When using more than 60K sentence pairs, we still get a relatively high promotion of translation quality. However, the promotion is not very remarkable as Row1-3 reveal in Table 4 . We can see when the number of parallel corpora is 100K(Row 5), the improvement over NMT Model is +3.95 BLEU points, which indicates that as the size of parallel corpora increases, the improvement of fine-tuning model is decreasing.",
                "cite_spans": [],
                "ref_spans": [
                    {
                        "start": 103,
                        "end": 110,
                        "text": "Table 4",
                        "ref_id": null
                    },
                    {
                        "start": 646,
                        "end": 653,
                        "text": "Table 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Corpus",
                "sec_num": null
            },
            {
                "text": "Most of existing work in neural machine translation focus on integrating SMT strategies Zhou et al., 2017; Wang et al., 2017; Shen et al., 2015) , handling rare words Sennrich et al., 2016b; Luong et al., 2015b) and designing the better framework (Tu et al., 2016; Luong et al., 2015a; Meng et al., 2016 Table 4 : Effect of different data size of parallel corpus. Method NMT Model means the result of conventional NMT system trained on these low-count parallel sentences. Partially Aligned Model(MSE) + Para means the result of our model fine-tuned by these parallel sentences. different scenarios has drawn intensive attention in recent years. Actually, there have been some effective methods to deal with them. We divide the related work into three categories:",
                "cite_spans": [
                    {
                        "start": 88,
                        "end": 106,
                        "text": "Zhou et al., 2017;",
                        "ref_id": "BIBREF34"
                    },
                    {
                        "start": 107,
                        "end": 125,
                        "text": "Wang et al., 2017;",
                        "ref_id": "BIBREF29"
                    },
                    {
                        "start": 126,
                        "end": 144,
                        "text": "Shen et al., 2015)",
                        "ref_id": "BIBREF26"
                    },
                    {
                        "start": 167,
                        "end": 190,
                        "text": "Sennrich et al., 2016b;",
                        "ref_id": "BIBREF25"
                    },
                    {
                        "start": 191,
                        "end": 211,
                        "text": "Luong et al., 2015b)",
                        "ref_id": "BIBREF19"
                    },
                    {
                        "start": 247,
                        "end": 264,
                        "text": "(Tu et al., 2016;",
                        "ref_id": "BIBREF28"
                    },
                    {
                        "start": 265,
                        "end": 285,
                        "text": "Luong et al., 2015a;",
                        "ref_id": "BIBREF18"
                    },
                    {
                        "start": 286,
                        "end": 303,
                        "text": "Meng et al., 2016",
                        "ref_id": "BIBREF20"
                    }
                ],
                "ref_spans": [
                    {
                        "start": 304,
                        "end": 311,
                        "text": "Table 4",
                        "ref_id": null
                    }
                ],
                "eq_spans": [],
                "section": "Related Work",
                "sec_num": "6"
            },
            {
                "text": "Pivot-based scenario assumes that there only exists source-pivot and pivot-target parallel corpora, which can be used to train source-to-pivot and pivot-to-target translation models. propose to translate source language into pivot language, and then the pivot language will be translated into target language. According to the fact that parallel sentences should have close probabilities of generating a sentence in a third language, Chen et al. (2017) construct a Teacher-Student framework, in which existing pivot-target NMT model guides the learning process of the source-target model.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Pivot-based Scenario",
                "sec_num": "6.1"
            },
            {
                "text": "In multilingual scenario, there exists multiple language pairs but no source-target sentence pairs. Johnson et al. (2016) use parallel corpora of multiple languages to train a universal NMT model. This universal model learns translation knowledge from multiple different languages, which makes zero-shot translation feasible. Firat et al. (2016) present a multi-way, multilingual model to resolve the zero-resource translation. They use other language to train a multi-way NMT model. The model generates pseudo parallel corpora to finetune attention mechanism, so as to achieve better translation.",
                "cite_spans": [
                    {
                        "start": 100,
                        "end": 121,
                        "text": "Johnson et al. (2016)",
                        "ref_id": "BIBREF10"
                    },
                    {
                        "start": 326,
                        "end": 345,
                        "text": "Firat et al. (2016)",
                        "ref_id": "BIBREF6"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Multilingual Scenario",
                "sec_num": "6.2"
            },
            {
                "text": "In this scenario, an NMT model of good quality has been trained on existing parallel corpora, but a preferable translation result is still in need by incorporating additional data resource. G\u00fcl\u00e7ehre et al. (2015) propose to incorporate target-side corpora as a language model. Sennrich et al. (2016a) attempt to enhance the decoder network model of NMT by incorporating the target-side monolingual data. Luong et at. (2016) explore the sequence autoencoders and skip-thought vectors method to exploit the monolingual data of source language. propose two approaches, self-training algorithm and multitask learning framework, to incorporate sourceside monolingual data. Besides that, Cheng et al. (2016) have explored the usage of both source and target monolingual data using a semisupervised method to reconstruct both source and target side monolingual language, where two NMT frameworks will be used.",
                "cite_spans": [
                    {
                        "start": 190,
                        "end": 212,
                        "text": "G\u00fcl\u00e7ehre et al. (2015)",
                        "ref_id": "BIBREF7"
                    },
                    {
                        "start": 277,
                        "end": 300,
                        "text": "Sennrich et al. (2016a)",
                        "ref_id": "BIBREF24"
                    },
                    {
                        "start": 682,
                        "end": 701,
                        "text": "Cheng et al. (2016)",
                        "ref_id": "BIBREF3"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Monolingual Data Scenario",
                "sec_num": "6.3"
            },
            {
                "text": "Above methods are designed for different scenarios, and their work can achieve great results on these scenarios. However, when in the scenario we propose in this work, that is we only have monolingual sentences and some phrase pairs, their methods are hard to be utilized to train an NMT model. Under this scenario, monolingual data can be acquired easily, and high quality phrase pairs can be obtained using some effective methods (Zhang et al., 2014) . To learn a good NMT model in our translation scenario, we adapt the conventional training procedure by designing a different generation mechanism and a different objective function.",
                "cite_spans": [
                    {
                        "start": 432,
                        "end": 452,
                        "text": "(Zhang et al., 2014)",
                        "ref_id": "BIBREF32"
                    }
                ],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Monolingual Data Scenario",
                "sec_num": "6.3"
            },
            {
                "text": "In this paper, we have presented a new translation scenario for NMT in which we have only monolingual data and bilingual phrase pairs. We obtain large-scale partially aligned sentence pairs from the monolingual data and phrase pairs by an information retrieval algorithm. The generation process and objective function are specially designed in NMT training to take full advantage of the partially aligned corpora. The empirical experiments show that the proposed method is capable to achieve a relatively good result. We further find that only a little amount of parallel sentences can significantly boost the translation quality.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "7"
            },
            {
                "text": "We also notice that the proposed approach with only partially aligned data cannot obtain high translation quality. In the future, we plan to design better approaches to model the partially aligned corpus. We also attempt to evaluate our approach on other language pairs, especially low-resource language pairs.",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Conclusion",
                "sec_num": "7"
            },
            {
                "text": "https://github.com/isi-nlp/ZophRNN",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "",
                "sec_num": null
            }
        ],
        "back_matter": [
            {
                "text": "The research work has been funded by the Natural Science Foundation of China under Grant No. 61333018 and No. 61402478, and it is also supported by the Strategic Priority Research Program of the CAS under Grant No. XDB02070007",
                "cite_spans": [],
                "ref_spans": [],
                "eq_spans": [],
                "section": "Acknowledgments",
                "sec_num": null
            }
        ],
        "bib_entries": {
            "BIBREF0": {
                "ref_id": "b0",
                "title": "Neural machine translation by jointly learning to align and translate",
                "authors": [
                    {
                        "first": "Dzmitry",
                        "middle": [],
                        "last": "Bahdanau",
                        "suffix": ""
                    },
                    {
                        "first": "Kyunghyun",
                        "middle": [],
                        "last": "Cho",
                        "suffix": ""
                    },
                    {
                        "first": "Yoshua",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of ICLR",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Dzmitry Bahdanau, Kyunghyun Cho, and Yoshua Ben- gio. 2015. Neural machine translation by jointly learning to align and translate. In Proceedings of ICLR 2015.",
                "links": null
            },
            "BIBREF1": {
                "ref_id": "b1",
                "title": "A teacher-student framework for zeroresource neural machine translation",
                "authors": [
                    {
                        "first": "Yun",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Yong",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Li",
                        "middle": [],
                        "last": "Victor",
                        "suffix": ""
                    },
                    {
                        "first": "O",
                        "middle": [
                            "K"
                        ],
                        "last": "",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yun Chen, Yong Cheng, Yang Liu, and Li Victor, O.K. 2017. A teacher-student framework for zero- resource neural machine translation. In Proceedings of ACL 2017.",
                "links": null
            },
            "BIBREF2": {
                "ref_id": "b2",
                "title": "Joint training for pivot-based neural machine translation",
                "authors": [
                    {
                        "first": "Yong",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Qian",
                        "middle": [],
                        "last": "Yang",
                        "suffix": ""
                    },
                    {
                        "first": "Maosong",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of IJCAI",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yong Cheng, Yang Liu, Qian Yang, Maosong Sun, and Wei Xu. 2017. Joint training for pivot-based neural machine translation. In Proceedings of IJCAI 2017.",
                "links": null
            },
            "BIBREF3": {
                "ref_id": "b3",
                "title": "Semisupervised learning for neural machine translation",
                "authors": [
                    {
                        "first": "Yong",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Zhongjun",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Hua",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Maosong",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yong Cheng, Wei Xu, Zhongjun He, Wei He, Hua Wu, Maosong Sun, and Yang Liu. 2016. Semi- supervised learning for neural machine translation. In Proceedings of ACL 2016.",
                "links": null
            },
            "BIBREF4": {
                "ref_id": "b4",
                "title": "A hierarchical phrase-based model for statistical machine translation",
                "authors": [
                    {
                        "first": "David",
                        "middle": [],
                        "last": "Chiang",
                        "suffix": ""
                    }
                ],
                "year": 2005,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "David Chiang. 2005. A hierarchical phrase-based model for statistical machine translation. In Pro- ceedings of ACL 2005.",
                "links": null
            },
            "BIBREF5": {
                "ref_id": "b5",
                "title": "Learning phrase representations using rnn encoder-decoder for statistical machine translation",
                "authors": [
                    {
                        "first": "Kyunghyun",
                        "middle": [],
                        "last": "Cho",
                        "suffix": ""
                    },
                    {
                        "first": "Bart",
                        "middle": [],
                        "last": "Van Merri\u00ebnboer Caglar Gulcehre",
                        "suffix": ""
                    },
                    {
                        "first": "Dzmitry",
                        "middle": [],
                        "last": "Bahdanau",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kyunghyun Cho, Bart van Merri\u00ebnboer Caglar Gul- cehre, Dzmitry Bahdanau, Fethi Bougares Holger Schwenk, and Yoshua Bengio. 2014. Learning phrase representations using rnn encoder-decoder for statistical machine translation. In Proceedings of EMNLP 2014.",
                "links": null
            },
            "BIBREF6": {
                "ref_id": "b6",
                "title": "Zero-resource translation with multi-lingual neural machine translation",
                "authors": [
                    {
                        "first": "Orhan",
                        "middle": [],
                        "last": "Firat",
                        "suffix": ""
                    },
                    {
                        "first": "Baskaran",
                        "middle": [],
                        "last": "Sankaran",
                        "suffix": ""
                    },
                    {
                        "first": "Yaser",
                        "middle": [],
                        "last": "Alonaizan",
                        "suffix": ""
                    },
                    {
                        "first": "T",
                        "middle": [
                            "Yarman"
                        ],
                        "last": "Fatos",
                        "suffix": ""
                    },
                    {
                        "first": "Kyunghyun",
                        "middle": [],
                        "last": "Vural",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Cho",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Orhan Firat, Baskaran Sankaran, Yaser Alonaizan, Fatos T. Yarman Vural, and Kyunghyun Cho. 2016. Zero-resource translation with multi-lingual neural machine translation. In Proceedings of EMNLP 2016.",
                "links": null
            },
            "BIBREF7": {
                "ref_id": "b7",
                "title": "On using monolingual corpora in neural machine translation",
                "authors": [
                    {
                        "first": "Caglar",
                        "middle": [],
                        "last": "G\u00fcl\u00e7ehre",
                        "suffix": ""
                    },
                    {
                        "first": "Orhan",
                        "middle": [],
                        "last": "Firat",
                        "suffix": ""
                    },
                    {
                        "first": "Kelvin",
                        "middle": [],
                        "last": "Xu",
                        "suffix": ""
                    },
                    {
                        "first": "Kyunghyun",
                        "middle": [],
                        "last": "Cho",
                        "suffix": ""
                    },
                    {
                        "first": "Lo\u0131c",
                        "middle": [],
                        "last": "Barrault",
                        "suffix": ""
                    },
                    {
                        "first": "Huei-Chi",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    },
                    {
                        "first": "Fethi",
                        "middle": [],
                        "last": "Bougares",
                        "suffix": ""
                    },
                    {
                        "first": "Holger",
                        "middle": [],
                        "last": "Schwenk",
                        "suffix": ""
                    },
                    {
                        "first": "Yoshua",
                        "middle": [],
                        "last": "Bengio",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Caglar G\u00fcl\u00e7ehre, Orhan Firat, Kelvin Xu, Kyunghyun Cho, Lo\u0131c Barrault, Huei-Chi Lin, Fethi Bougares, Holger Schwenk, and Yoshua Bengio. 2015. On us- ing monolingual corpora in neural machine transla- tion. CoRR, abs/1503.03535.",
                "links": null
            },
            "BIBREF8": {
                "ref_id": "b8",
                "title": "Improved neural machine translation with smt features",
                "authors": [
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Zhongjun",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Hua",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Haifeng",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of AAAI",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Wei He, Zhongjun He, Hua Wu, and Haifeng Wang. 2016. Improved neural machine translation with smt features. In Proceedings of AAAI 2016.",
                "links": null
            },
            "BIBREF9": {
                "ref_id": "b9",
                "title": "Long short-term memory",
                "authors": [
                    {
                        "first": "Sepp",
                        "middle": [],
                        "last": "Hochreiter",
                        "suffix": ""
                    },
                    {
                        "first": "J\u00fcrgen",
                        "middle": [],
                        "last": "Schmidhuber",
                        "suffix": ""
                    }
                ],
                "year": 1997,
                "venue": "Neural computation",
                "volume": "9",
                "issue": "8",
                "pages": "1735--1780",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Sepp Hochreiter and J\u00fcrgen Schmidhuber. 1997. Long short-term memory. Neural computation, 9(8):1735-1780.",
                "links": null
            },
            "BIBREF10": {
                "ref_id": "b10",
                "title": "Google's multilingual neural machine translation system: Enabling zero-shot translation",
                "authors": [
                    {
                        "first": "Melvin",
                        "middle": [],
                        "last": "Johnson",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Schuster",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Quoc",
                        "suffix": ""
                    },
                    {
                        "first": "Maxim",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Yonghui",
                        "middle": [],
                        "last": "Krikun",
                        "suffix": ""
                    },
                    {
                        "first": "Zhifeng",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Nikhil",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "Fernanda",
                        "middle": [],
                        "last": "Thorat",
                        "suffix": ""
                    },
                    {
                        "first": "Martin",
                        "middle": [],
                        "last": "Vi\u00e9gas",
                        "suffix": ""
                    },
                    {
                        "first": "Greg",
                        "middle": [],
                        "last": "Wattenberg",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Corrado",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1611.04558"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Melvin Johnson, Mike Schuster, Quoc V Le, Maxim Krikun, Yonghui Wu, Zhifeng Chen, Nikhil Tho- rat, Fernanda Vi\u00e9gas, Martin Wattenberg, and Greg Corrado. 2016. Google's multilingual neural ma- chine translation system: Enabling zero-shot trans- lation. arXiv preprint arXiv:1611.04558.",
                "links": null
            },
            "BIBREF11": {
                "ref_id": "b11",
                "title": "Recurrent continuous translation models",
                "authors": [
                    {
                        "first": "Nal",
                        "middle": [],
                        "last": "Kalchbrenner",
                        "suffix": ""
                    },
                    {
                        "first": "Phil",
                        "middle": [],
                        "last": "Blunsom",
                        "suffix": ""
                    }
                ],
                "year": 2013,
                "venue": "Proceedings of EMNLP 2013",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Nal Kalchbrenner and Phil Blunsom. 2013. Recurrent continuous translation models. In Proceedings of EMNLP 2013.",
                "links": null
            },
            "BIBREF12": {
                "ref_id": "b12",
                "title": "Moses: Open source toolkit for statistical machine translation",
                "authors": [
                    {
                        "first": "Philipp",
                        "middle": [],
                        "last": "Koehn",
                        "suffix": ""
                    },
                    {
                        "first": "Hieu",
                        "middle": [],
                        "last": "Hoang",
                        "suffix": ""
                    },
                    {
                        "first": "Alexandra",
                        "middle": [],
                        "last": "Birch",
                        "suffix": ""
                    },
                    {
                        "first": "Chris",
                        "middle": [],
                        "last": "Callison-Burch",
                        "suffix": ""
                    },
                    {
                        "first": "Marcello",
                        "middle": [],
                        "last": "Federico",
                        "suffix": ""
                    },
                    {
                        "first": "Nicola",
                        "middle": [],
                        "last": "Bertoldi",
                        "suffix": ""
                    },
                    {
                        "first": "Brooke",
                        "middle": [],
                        "last": "Cowan",
                        "suffix": ""
                    },
                    {
                        "first": "Wade",
                        "middle": [],
                        "last": "Shen",
                        "suffix": ""
                    },
                    {
                        "first": "Christine",
                        "middle": [],
                        "last": "Moran",
                        "suffix": ""
                    },
                    {
                        "first": "Richard",
                        "middle": [],
                        "last": "Zens",
                        "suffix": ""
                    }
                ],
                "year": 2007,
                "venue": "Proceedings of ACL 2007",
                "volume": "",
                "issue": "",
                "pages": "177--180",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Philipp Koehn, Hieu Hoang, Alexandra Birch, Chris Callison-Burch, Marcello Federico, Nicola Bertoldi, Brooke Cowan, Wade Shen, Christine Moran, Richard Zens, et al. 2007. Moses: Open source toolkit for statistical machine translation. In Pro- ceedings of ACL 2007, pages 177-180.",
                "links": null
            },
            "BIBREF13": {
                "ref_id": "b13",
                "title": "Statistical phrase-based translation",
                "authors": [
                    {
                        "first": "Philipp",
                        "middle": [],
                        "last": "Koehn",
                        "suffix": ""
                    },
                    {
                        "first": "Franz",
                        "middle": [
                            "J"
                        ],
                        "last": "Och",
                        "suffix": ""
                    },
                    {
                        "first": "Daniel",
                        "middle": [],
                        "last": "Marcu",
                        "suffix": ""
                    }
                ],
                "year": 2003,
                "venue": "Proceedings of ACL-NAACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Philipp Koehn, Franz J. Och, and Daniel Marcu. 2003. Statistical phrase-based translation. In Proceedings of ACL-NAACL 2013.",
                "links": null
            },
            "BIBREF14": {
                "ref_id": "b14",
                "title": "Towards zero unknown word in neural machine translation",
                "authors": [
                    {
                        "first": "Xiaoqing",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Jiajun",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Chengqing",
                        "middle": [],
                        "last": "Zong",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of IJCAI",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xiaoqing Li, Jiajun Zhang, and Chengqing Zong. 2016. Towards zero unknown word in neural machine translation. In Proceedings of IJCAI 2016.",
                "links": null
            },
            "BIBREF15": {
                "ref_id": "b15",
                "title": "Neural machine translation with supervised attention",
                "authors": [
                    {
                        "first": "Lemao",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Masao",
                        "middle": [],
                        "last": "Utiyama",
                        "suffix": ""
                    },
                    {
                        "first": "Andrew",
                        "middle": [],
                        "last": "Finch",
                        "suffix": ""
                    },
                    {
                        "first": "Eiichiro",
                        "middle": [],
                        "last": "Sumita",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of COL-ING",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Lemao Liu, Masao Utiyama, Andrew Finch, and Ei- ichiro Sumita. 2016. Neural machine translation with supervised attention. In Proceedings of COL- ING 2016.",
                "links": null
            },
            "BIBREF16": {
                "ref_id": "b16",
                "title": "Treeto-string alignment template for statistical machine translation",
                "authors": [
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Qun",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Shouxun",
                        "middle": [],
                        "last": "Lin",
                        "suffix": ""
                    }
                ],
                "year": 2006,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Yang Liu, Qun Liu, and Shouxun Lin. 2006. Tree- to-string alignment template for statistical machine translation. In Proceedings of ACL 2006.",
                "links": null
            },
            "BIBREF17": {
                "ref_id": "b17",
                "title": "Multi-task sequence to sequence learning",
                "authors": [
                    {
                        "first": "Minh-Thang",
                        "middle": [],
                        "last": "Luong",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Quoc",
                        "suffix": ""
                    },
                    {
                        "first": "Ilya",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Oriol",
                        "middle": [],
                        "last": "Sutskever",
                        "suffix": ""
                    },
                    {
                        "first": "Lukasz",
                        "middle": [],
                        "last": "Vinyals",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Kaiser",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of ICLR",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Minh-Thang Luong, Quoc V Le, Ilya Sutskever, Oriol Vinyals, and Lukasz Kaiser. 2016. Multi-task se- quence to sequence learning. In Proceedings of ICLR 2016.",
                "links": null
            },
            "BIBREF18": {
                "ref_id": "b18",
                "title": "Effective approaches to attentionbased neural machine translation",
                "authors": [
                    {
                        "first": "Minh-Thang",
                        "middle": [],
                        "last": "Luong",
                        "suffix": ""
                    },
                    {
                        "first": "Hieu",
                        "middle": [],
                        "last": "Pham",
                        "suffix": ""
                    },
                    {
                        "first": "Christopher D",
                        "middle": [],
                        "last": "Manning",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Minh-Thang Luong, Hieu Pham, and Christopher D Manning. 2015a. Effective approaches to attention- based neural machine translation. In Proceedings of EMNLP 2015.",
                "links": null
            },
            "BIBREF19": {
                "ref_id": "b19",
                "title": "Addressing the rare word problem in neural machine translation",
                "authors": [
                    {
                        "first": "Minh-Thang",
                        "middle": [],
                        "last": "Luong",
                        "suffix": ""
                    },
                    {
                        "first": "Ilya",
                        "middle": [],
                        "last": "Sutskever",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Quoc",
                        "suffix": ""
                    },
                    {
                        "first": "Oriol",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Wojciech",
                        "middle": [],
                        "last": "Vinyals",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Zaremba",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Minh-Thang Luong, Ilya Sutskever, Quoc V Le, Oriol Vinyals, and Wojciech Zaremba. 2015b. Addressing the rare word problem in neural machine translation. In Proceedings of ACL 2015.",
                "links": null
            },
            "BIBREF20": {
                "ref_id": "b20",
                "title": "Interactive attention for neural machine translation",
                "authors": [
                    {
                        "first": "Fandong",
                        "middle": [],
                        "last": "Meng",
                        "suffix": ""
                    },
                    {
                        "first": "Zhengdong",
                        "middle": [],
                        "last": "Lu",
                        "suffix": ""
                    },
                    {
                        "first": "Hang",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Qun",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of COLING",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Fandong Meng, Zhengdong Lu, Hang Li, and Qun Liu. 2016. Interactive attention for neural machine trans- lation. In In Proceedings of COLING 2016.",
                "links": null
            },
            "BIBREF21": {
                "ref_id": "b21",
                "title": "Supervised attentions for neural machine translation",
                "authors": [
                    {
                        "first": "Haitao",
                        "middle": [],
                        "last": "Mi",
                        "suffix": ""
                    },
                    {
                        "first": "Zhiguo",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Abe",
                        "middle": [],
                        "last": "Ittycheriah",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Haitao Mi, Zhiguo Wang, and Abe Ittycheriah. 2016a. Supervised attentions for neural machine translation. In Proceedings of EMNLP 2016.",
                "links": null
            },
            "BIBREF22": {
                "ref_id": "b22",
                "title": "Vocabulary manipulation for neural machine translation",
                "authors": [
                    {
                        "first": "Haitao",
                        "middle": [],
                        "last": "Mi",
                        "suffix": ""
                    },
                    {
                        "first": "Zhiguo",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Abe",
                        "middle": [],
                        "last": "Ittycheriah",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Haitao Mi, Zhiguo Wang, and Abe Ittycheriah. 2016b. Vocabulary manipulation for neural machine trans- lation. In Proceedings of ACL 2016.",
                "links": null
            },
            "BIBREF23": {
                "ref_id": "b23",
                "title": "Bleu: a method for automatic evaluation of machine translation",
                "authors": [
                    {
                        "first": "Kishore",
                        "middle": [],
                        "last": "Papineni",
                        "suffix": ""
                    },
                    {
                        "first": "Salim",
                        "middle": [],
                        "last": "Roukos",
                        "suffix": ""
                    },
                    {
                        "first": "Todd",
                        "middle": [],
                        "last": "Ward",
                        "suffix": ""
                    },
                    {
                        "first": "Wei-Jing",
                        "middle": [],
                        "last": "Zhu",
                        "suffix": ""
                    }
                ],
                "year": 2002,
                "venue": "Proceedings of ACL 2002",
                "volume": "",
                "issue": "",
                "pages": "311--318",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Kishore Papineni, Salim Roukos, Todd Ward, and Wei- Jing Zhu. 2002. Bleu: a method for automatic evalu- ation of machine translation. In Proceedings of ACL 2002, pages 311-318.",
                "links": null
            },
            "BIBREF24": {
                "ref_id": "b24",
                "title": "Improving neural machine translation models with monolingual data",
                "authors": [
                    {
                        "first": "Rico",
                        "middle": [],
                        "last": "Sennrich",
                        "suffix": ""
                    },
                    {
                        "first": "Barry",
                        "middle": [],
                        "last": "Haddow",
                        "suffix": ""
                    },
                    {
                        "first": "Alexandra",
                        "middle": [],
                        "last": "Birch",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016a. Improving neural machine translation mod- els with monolingual data. In Proceedings of ACL 2016.",
                "links": null
            },
            "BIBREF25": {
                "ref_id": "b25",
                "title": "Neural machine translation of rare words with subword units",
                "authors": [
                    {
                        "first": "Rico",
                        "middle": [],
                        "last": "Sennrich",
                        "suffix": ""
                    },
                    {
                        "first": "Barry",
                        "middle": [],
                        "last": "Haddow",
                        "suffix": ""
                    },
                    {
                        "first": "Alexandra",
                        "middle": [],
                        "last": "Birch",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Rico Sennrich, Barry Haddow, and Alexandra Birch. 2016b. Neural machine translation of rare words with subword units. In Proceedings of ACL 2016.",
                "links": null
            },
            "BIBREF26": {
                "ref_id": "b26",
                "title": "Minimum risk training for neural machine translation",
                "authors": [
                    {
                        "first": "Shiqi",
                        "middle": [],
                        "last": "Shen",
                        "suffix": ""
                    },
                    {
                        "first": "Yong",
                        "middle": [],
                        "last": "Cheng",
                        "suffix": ""
                    },
                    {
                        "first": "Zhongjun",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Wei",
                        "middle": [],
                        "last": "He",
                        "suffix": ""
                    },
                    {
                        "first": "Hua",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Maosong",
                        "middle": [],
                        "last": "Sun",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    }
                ],
                "year": 2015,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Shiqi Shen, Yong Cheng, Zhongjun He, Wei He, Hua Wu, Maosong Sun, and Yang Liu. 2015. Minimum risk training for neural machine translation. In Pro- ceedings of ACL 2015.",
                "links": null
            },
            "BIBREF27": {
                "ref_id": "b27",
                "title": "Sequence to sequence learning with neural networks",
                "authors": [
                    {
                        "first": "Ilya",
                        "middle": [],
                        "last": "Sutskever",
                        "suffix": ""
                    },
                    {
                        "first": "Oriol",
                        "middle": [],
                        "last": "Vinyals",
                        "suffix": ""
                    },
                    {
                        "first": "Quoc V",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of NIPS",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Ilya Sutskever, Oriol Vinyals, and Quoc V Le. 2014. Sequence to sequence learning with neural net- works. In Proceedings of NIPS 2014.",
                "links": null
            },
            "BIBREF28": {
                "ref_id": "b28",
                "title": "Coverage-based neural machine translation",
                "authors": [
                    {
                        "first": "Zhaopeng",
                        "middle": [],
                        "last": "Tu",
                        "suffix": ""
                    },
                    {
                        "first": "Zhengdong",
                        "middle": [],
                        "last": "Lu",
                        "suffix": ""
                    },
                    {
                        "first": "Yang",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Xiaohua",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Hang",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Zhaopeng Tu, Zhengdong Lu, Yang Liu, Xiaohua Liu, and Hang Li. 2016. Coverage-based neural machine translation. In Proceedings of ACL 2016.",
                "links": null
            },
            "BIBREF29": {
                "ref_id": "b29",
                "title": "Neural machine translation advised by statistical machine translation",
                "authors": [
                    {
                        "first": "Xing",
                        "middle": [],
                        "last": "Wang",
                        "suffix": ""
                    },
                    {
                        "first": "Zhengdong",
                        "middle": [],
                        "last": "Lu",
                        "suffix": ""
                    },
                    {
                        "first": "Zhaopeng",
                        "middle": [],
                        "last": "Tu",
                        "suffix": ""
                    },
                    {
                        "first": "Hang",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Deyi",
                        "middle": [],
                        "last": "Xiong",
                        "suffix": ""
                    },
                    {
                        "first": "Min",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of AAAI",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Xing Wang, Zhengdong Lu, Zhaopeng Tu, Hang Li, Deyi Xiong, and Min Zhang. 2017. Neural machine translation advised by statistical machine transla- tion. In Proceedings of AAAI 2017.",
                "links": null
            },
            "BIBREF30": {
                "ref_id": "b30",
                "title": "Google's neural machine translation system: Bridging the gap between human and machine translation",
                "authors": [
                    {
                        "first": "Yonghui",
                        "middle": [],
                        "last": "Wu",
                        "suffix": ""
                    },
                    {
                        "first": "Mike",
                        "middle": [],
                        "last": "Schuster",
                        "suffix": ""
                    },
                    {
                        "first": "Zhifeng",
                        "middle": [],
                        "last": "Chen",
                        "suffix": ""
                    },
                    {
                        "first": "V",
                        "middle": [],
                        "last": "Quoc",
                        "suffix": ""
                    },
                    {
                        "first": "Mohammad",
                        "middle": [],
                        "last": "Le",
                        "suffix": ""
                    },
                    {
                        "first": "Wolfgang",
                        "middle": [],
                        "last": "Norouzi",
                        "suffix": ""
                    },
                    {
                        "first": "Maxim",
                        "middle": [],
                        "last": "Macherey",
                        "suffix": ""
                    },
                    {
                        "first": "Yuan",
                        "middle": [],
                        "last": "Krikun",
                        "suffix": ""
                    },
                    {
                        "first": "Qin",
                        "middle": [],
                        "last": "Cao",
                        "suffix": ""
                    },
                    {
                        "first": "Klaus",
                        "middle": [],
                        "last": "Gao",
                        "suffix": ""
                    },
                    {
                        "first": "",
                        "middle": [],
                        "last": "Macherey",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {
                    "arXiv": [
                        "arXiv:1609.08144"
                    ]
                },
                "num": null,
                "urls": [],
                "raw_text": "Yonghui Wu, Mike Schuster, Zhifeng Chen, Quoc V Le, Mohammad Norouzi, Wolfgang Macherey, Maxim Krikun, Yuan Cao, Qin Gao, Klaus Macherey, et al. 2016. Google's neural ma- chine translation system: Bridging the gap between human and machine translation. arXiv preprint arXiv:1609.08144.",
                "links": null
            },
            "BIBREF31": {
                "ref_id": "b31",
                "title": "Tree-based translation without using parse trees",
                "authors": [
                    {
                        "first": "Feifei",
                        "middle": [],
                        "last": "Zhai",
                        "suffix": ""
                    },
                    {
                        "first": "Jiajun",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Yu",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Chengqing",
                        "middle": [],
                        "last": "Zong",
                        "suffix": ""
                    }
                ],
                "year": 2012,
                "venue": "Proceedings of COLING",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Feifei Zhai, Jiajun Zhang, Yu Zhou, Chengqing Zong, et al. 2012. Tree-based translation without using parse trees. In Proceedings of COLING 2012.",
                "links": null
            },
            "BIBREF32": {
                "ref_id": "b32",
                "title": "Bilingually-constrained phrase embeddings for machine translation",
                "authors": [
                    {
                        "first": "Jiajun",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Shujie",
                        "middle": [],
                        "last": "Liu",
                        "suffix": ""
                    },
                    {
                        "first": "Mu",
                        "middle": [],
                        "last": "Li",
                        "suffix": ""
                    },
                    {
                        "first": "Ming",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Chengqing",
                        "middle": [],
                        "last": "Zong",
                        "suffix": ""
                    }
                ],
                "year": 2014,
                "venue": "Proceedings of ACL",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jiajun Zhang, Shujie Liu, Mu Li, Ming Zhou, and Chengqing Zong. 2014. Bilingually-constrained phrase embeddings for machine translation. In Pro- ceedings of ACL 2014.",
                "links": null
            },
            "BIBREF33": {
                "ref_id": "b33",
                "title": "Exploiting source-side monolingual data in neural machine translation",
                "authors": [
                    {
                        "first": "Jiajun",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Chengqing",
                        "middle": [],
                        "last": "Zong",
                        "suffix": ""
                    }
                ],
                "year": 2016,
                "venue": "Proceedings of EMNLP",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Jiajun Zhang and Chengqing Zong. 2016. Exploit- ing source-side monolingual data in neural machine translation. In Proceedings of EMNLP 2016.",
                "links": null
            },
            "BIBREF34": {
                "ref_id": "b34",
                "title": "Neural system combination for machine translation",
                "authors": [
                    {
                        "first": "Long",
                        "middle": [],
                        "last": "Zhou",
                        "suffix": ""
                    },
                    {
                        "first": "Wenpeng",
                        "middle": [],
                        "last": "Hu",
                        "suffix": ""
                    },
                    {
                        "first": "Jiajun",
                        "middle": [],
                        "last": "Zhang",
                        "suffix": ""
                    },
                    {
                        "first": "Chengqing",
                        "middle": [],
                        "last": "Zong",
                        "suffix": ""
                    }
                ],
                "year": 2017,
                "venue": "Proceedings of ACL 2017",
                "volume": "",
                "issue": "",
                "pages": "",
                "other_ids": {},
                "num": null,
                "urls": [],
                "raw_text": "Long Zhou, Wenpeng Hu, Jiajun Zhang, and Chengqing Zong. 2017. Neural system combina- tion for machine translation. In Proceedings of ACL 2017.",
                "links": null
            }
        },
        "ref_entries": {
            "FIGREF0": {
                "text": "The framework of our training method for partially aligned sentence, in which (x 2 ,x 3 ) and (y 2 ,y 3 ) are parallel parts.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "FIGREF1": {
                "text": "Translation results (BLEU score) of different lengths.",
                "type_str": "figure",
                "num": null,
                "uris": null
            },
            "TABREF0": {
                "type_str": "table",
                "num": null,
                "text": "",
                "content": "<table><tr><td>Corpus</td><td/><td>Chinese</td><td>English</td></tr><tr><td/><td colspan=\"2\">#Sent. 13.33M</td><td>10.03M</td></tr><tr><td>monolingual</td><td colspan=\"3\">#Word 327.10M 276.07M</td></tr><tr><td/><td>Vocab</td><td>1.83M</td><td>1.07M</td></tr><tr><td colspan=\"4\">Table 1: The statistics of monolingual dataset on</td></tr><tr><td>the LDC corpus.</td><td/><td/></tr></table>",
                "html": null
            },
            "TABREF2": {
                "type_str": "table",
                "num": null,
                "text": "Translation results (BLEU score) for different translation methods.",
                "content": "<table><tr><td>testing, we employ beam search algorithm, and the</td></tr><tr><td>beam size is 12.</td></tr></table>",
                "html": null
            },
            "TABREF3": {
                "type_str": "table",
                "num": null,
                "text": "Partially Aligned Model(MSE) + Para 12.36 15.07 11.64 14.61 13.42 Partially Aligned Model(MSE) + Para 14.12 17.84 13.66 17.26 15.72",
                "content": "<table><tr><td>#Sent.</td><td>Method</td><td colspan=\"5\">MT03 MT04 MT05 MT06 Ave</td></tr><tr><td>20K</td><td>NMT Model</td><td>1.60</td><td>1.22</td><td>1.05</td><td>1.70</td><td>1.39</td></tr><tr><td>40K</td><td>NMT Model</td><td>1.87</td><td>2.00</td><td>1.47</td><td>2.24</td><td>1.90</td></tr><tr><td>60K</td><td colspan=\"6\">NMT Model Partially Aligned Model(MSE) + Para 15.54 19.62 15.16 19.25 17.39 3.72 5.04 3.49 4.47 4.18</td></tr><tr><td>80K</td><td colspan=\"6\">NMT Model Partially Aligned Model(MSE) + Para 17.16 21.18 16.65 20.61 18.90 7.96 11.85 8.16 10.53 9.63</td></tr><tr><td>100K</td><td colspan=\"6\">NMT Model Partially Aligned(MSE) Model + Para 18.30 22.50 17.92 21.55 20.07 14.50 18.21 14.29 17.49 16.12</td></tr><tr><td/><td/><td/><td/><td/><td/><td>). As for</td></tr><tr><td/><td/><td colspan=\"5\">translation scenarios, training NMT model under</td></tr></table>",
                "html": null
            }
        }
    }
}