Skip to main content

Über dieses Buch

This book constitutes the refereed proceedings of the 13th China Workshop on Machine Translation, CWMT 2017, held in Dalian, China, in September 2017.
The 10 papers presented in this volume were carefully reviewed and selected from 26 submissions and focus on all aspects of machine translation, including preprocessing, neural machine translation models, hybrid model, evaluation method, and post-editing.



Neural Machine Translation with Phrasal Attention

Attention-based neural machine translation (NMT) employs an attention network to capture structural correspondences between the source and target language at the word level. Unfortunately, alignments between source and target equivalents are complicated, which makes word-level attention not adequate to model these relations (e.g., alignments between a source idiom and its target translation). In order to handle this issue, we propose a phrase-level attention mechanism to complement the word-level attention network in this paper. The proposed phrasal attention framework is simple yet effective, keeping the strength of phrase-based statistical machine translation (SMT) on the source side. Experiments on Chinese-to-English translation task demonstrate that the proposed method is able to statistically improve word-level attention-based NMT.

Yachao Li, Deyi Xiong, Min Zhang

Singleton Detection for Coreference Resolution via Multi-window and Multi-filter CNN

Mention detection is the first and a key stage in most of coreference resolution systems. Singleton mentions are the ones which appear only once and are not mentioned in the following texts. Singleton mentions always affect the performance of coreference resolution systems. To remove the singleton ones from the automatically predicted mentions, we propose a novel singleton detection method based on multi-window and multi-filter convolutional neural network (MMCNN). The MMCNN model can detect singleton mentions with less use of hand-designed features and more sentence information. Experiments show that our system outperforms all the existing singleton detection systems.

Kenan Li, Heyan Huang, Yuhang Guo, Ping Jian

A Method of Unknown Words Processing for Neural Machine Translation Using HowNet

An inherent weakness of neural machine translation (NMT) systems is their inability to correctly translate unknown words. Traditional unknown words processing methods are usually based on word vectors trained on large scale of monolingual corpus. Replacing the unknown words according to the similarity of word vectors. However, it suffers from two weaknesses: Firstly, the resulting vectors of unknown words are not of high quality; Secondly, it is difficult to deal with polysemous words. This paper proposes an unknown word processing method by integrating HowNet. Using the concepts and sememes in HowNet to seek the replacement words of unknown words. Experimental results show that our proposed method can not only improves the performance of NMT, but also provides some advantages compared with the traditional unknown words processing methods.

Shaotong Li, JinAn Xu, Yujie Zhang, Yufeng Chen

Word, Subword or Character? An Empirical Study of Granularity in Chinese-English NMT

Neural machine translation (NMT) becomes a new approach to machine translation and is proved to outperform conventional statistical machine translation (SMT) across a variety of language pairs. Most existing NMT systems operate with a fixed vocabulary, but translation is an open-vocabulary problem. Hence, previous works mainly handle rare and unknown words by using different translation granularities, such as character, subword, and hybrid word-character. While translation involving Chinese has been proved to be one of the most difficult tasks, there is no study to demonstrate which translation granularity is the most suitable for Chinese in NMT. In this paper, we conduct an extensive comparison using Chinese-English NMT as a case study. Furthermore, we discuss the advantages and disadvantages of various translation granularities in detail. Our experiments show that subword model performs best for Chinese-to-English translation while hybrid word-character model is most suitable for English-to-Chinese translation.

Yining Wang, Long Zhou, Jiajun Zhang, Chengqing Zong

An Unknown Word Processing Method in NMT by Integrating Syntactic Structure and Semantic Concept

The unknown words in neural machine translation (NMT) may undermine the integrity of sentence structure, increase ambiguity and have adverse effect on the translation. In order to solve this problem, we propose a method of processing unknown words in NMT based on integrating syntactic structure and semantic concept. Firstly, the semantic concept network is used to construct the set of in-vocabulary synonyms corresponding to the unknown words. Secondly, a semantic similarity calculation method based on the syntactic structure and semantic concept is proposed. The best substitute is selected from the set of in-vocabulary synonyms by calculating the semantic similarity between the unknown words and their candidate substitutes. English-Chinese translation experiments demonstrate that this method can maintain the semantic integrity of the source language sentences. Meanwhile, in performance, our proposed method can obtain an improvement by 2.9 BLEU points when compared with the conventional NMT method, and the method can also achieve an improvement by 0.95 BLEU points when compared with the traditional method of positioning the UNK character based on word alignment information.

Guoyi Miao, Jinan Xu, Yancui Li, Shaotong Li, Yufeng Chen

RGraph: Generating Reference Graphs for Better Machine Translation Evaluation

Statistical machine translation systems perform parameter learning (i.e. training) basing on automatic translation evaluation methods, which usually evaluate the translation quality according to one or more human-translated references. Although producing more references would improve the coverage of translation choices and lead to improved training performances, only several references are used due to the cost of human translation. In this paper, we propose automatic methods to explore the information among the limited references. By generating a reference graph (RGraph) from given references, we could automatically generate exponential number of references. These diverse references make it possible to better evaluate each individual translations, without using any other resources. Experiments showed that our RGraph could improve the evaluation performance and lead to better tuned machine translation systems. The method could be extended to improve the evaluation with single reference as well.

Hongjie Ji, Shujian Huang, Qi Hou, Cunyan Yin, Jiajun Chen

ENTF: An Entropy-Based MT Evaluation Metric

The widely-used automatic evaluation metrics cannot adequately reflect the fluency of the translations. The n-gram-based metrics, like BLEU, limit the maximum length of matched fragments to n and cannot catch the matched fragments longer than n, so they can only reflect the fluency indirectly. METEOR, which is not limited by n-gram, uses the number of matched chunks but it does not consider the length of each chunk. In this paper, we propose an entropy-based metric (ENTF), which can sufficiently reflect the fluency of translations through the distribution of matched words. To evaluate the accuracy, we also introduce the unigram F-score into the new metric. Experiment shows that ENTF obtains state-of-the-art performance on system level, and is comparable with METEOR on sentence level on into English direction on WMT 2012, WMT 2013 and WMT 2014.

Hui Yu, Weizhi Xu, Shouxun Lin, Qun Liu

Translation Oriented Sentence Level Collocation Identification and Extraction

The technique to identify and extract collocations in a given sentence is very important to sentence understanding, analysing and translating. So we propose a sentence level collocation identification and extraction method which follows the traditional two phase collocation extraction model. In candidate generating phase, we use the dependency parsing results directly, while in the filtering phase, we propose to use the latest model of distributional semantics - word embedding based similarity to filter the noises. For each candidate, three word embedding based similarity rankings will be obtained and accordingly to decide if it is a real collocation. The experimental results show that the proposed filtering method performs better than the traditional well-known association measures. The comparison with the baseline system shows that the proposed method can retrieve more collocations with higher precision than the baseline, which is of significance to sentence related natural language processing tasks.

Xiaoxia Liu, Degen Huang

Combining Domain Knowledge and Deep Learning Makes NMT More Adaptive

In both SMT (statistical machine translation) and NMT (neural machine translation), training data often varies in source, theme and genre. It is less likely that the training data and texts in practical translation fall into a same domain, leading to a sub-optimal performance. Domain adaptation is to address such problems. Existing domain adaptive approach in machine translation employs topic model to obtain topic information. However, thus domain labels can be very much limited to in-domain and out-of-domain, when dividing topics into two types, without any more specific labels. We propose a novel domain adaptive approach to annotate Chinese sentences with CLCN (Chinese Library Classification Number) as the domain labels. We design a deep fusion model of neural network to combine two annotating models, including one applying a domain knowledge base built on thesis keywords and Chinese Scientific and Technical Vocabulary System, and the other applying deep learning method based on a CNN. Then, we have the fused domain annotator to filter the training data of NMT according to the test data. After running two predefined domain test sets on a NMT system trained by only partial of the original training data, we achieve an average 1.3 BLEU score improvement (5.4% relative), which demonstrates the feasibility and validity of proposed approach.

Liang Ding, Yanqing He, Lei Zhou, Qingmin Liu

Handling Many-To-One UNK Translation for Neural Machine Translation

Neural machine translation has achieved remarkable progress recently, but it is restricted by a limited vocabulary due to the computation complexity. All words out of the vocabulary are replaced with a single UNK, and the UNK in translation results will hurt the quality of translation. In this paper, a UNK translation method is proposed to handle the unknown word issue in neural machine translation. It uses n-best source alignment candidates for UNK translation, and can handle both word level (one-to-one) and phrase level (many-to-one) source-UNK alignment. Experiments on Chinese-to-English task shows that our method achieves a +0.73 BLEU improvement over the NMT baseline that has already employed a good UNK translation module.

Fuxue Li, Du Quan, Wang Qiang, Xiao Tong, Jingbo Zhu

A Content-Based Neural Reordering Model for Statistical Machine Translation

Phrase-based lexicalized reordering models have attracted extensive interest in statistical machine translation (SMT) due to their capacity for dealing with swap between consecutive phrases. However, translations between two languages that with significant differences in syntactic structure have made it challenging to generate a semantically and syntactically correct word sequence. In an effort to alleviate this problem, we propose a novel content-based neural reordering model that estimates reordering probabilities based on the words of its surrounding contexts. We first utilize a simple convolutional neural network (CNN) to capture semantic contents conditioned on various sizes of context. And then we employ a softmax layer to predict the reordering orientations and probability distributions. Experimental results show that our model provides statistically obvious improvements for both Chinese-Uyghur (+0.48 on CWMT2015) and Chinese-English (+0.27 on CWMT2013) translation tasks over conventional lexicalized reordering models.

Yirong Pan, Xiao Li, Yating Yang, Chenggang Mi, Rui Dong, Wenxiao Zeng


Weitere Informationen

Premium Partner

Neuer Inhalt

BranchenIndex Online

Die B2B-Firmensuche für Industrie und Wirtschaft: Kostenfrei in Firmenprofilen nach Lieferanten, Herstellern, Dienstleistern und Händlern recherchieren.



Product Lifecycle Management im Konzernumfeld – Herausforderungen, Lösungsansätze und Handlungsempfehlungen

Für produzierende Unternehmen hat sich Product Lifecycle Management in den letzten Jahrzehnten in wachsendem Maße zu einem strategisch wichtigen Ansatz entwickelt. Forciert durch steigende Effektivitäts- und Effizienzanforderungen stellen viele Unternehmen ihre Product Lifecycle Management-Prozesse und -Informationssysteme auf den Prüfstand. Der vorliegende Beitrag beschreibt entlang eines etablierten Analyseframeworks Herausforderungen und Lösungsansätze im Product Lifecycle Management im Konzernumfeld.
Jetzt gratis downloaden!