Can the Translation Memory Principle Benefit Neural Machine Translation? A Series of Extensive Experiments with Input Sentence Annotation

Can the Translation Memory Principle Benefit Neural Machine Translation? A Series of Extensive Experiments with Input Sentence Annotation Yaling Wang Graduate School of IPS Waseda University Kitakyushu, Japan yaling.wang@moegi.waseda.jp Yves Lepage Graduate School of IPS Waseda University Kitakyushu, Japan yves.lepage@waseda.jp Abstract Integrating translation memories (TM) into neural machine translation (NMT) has been shown to improve translation quality. We test various schemes to integrate translation sug- gestions into an NMT system without alter- ing its architecture. We retrieve similar sen- tences covering the sentence to translate and examine various annotation schemes as input to the NMT system. Our results show that the method can outperform a baseline model in some cases. The improvements are mainly for the translation of sentences with a length ranging from 10 to 20 words. 1 Introduction Translation Memories (TMs) are used daily by translators. They contain aligned parallel sentence pairs. Given a sentence to translate, a TM retrieves the most similar sentence in the source language that contains large common or similar parts. The corre- sponding sentence in the target language is returned to the translator. In this way, the translator needs only to modify the unmatched parts to complete the translation. A main advantage of translation memo- ries is that they ensure consistency and interpretabil- ity across translations because common or similar parts in sentences can easily be identified. Recently, with the development of neural machine translation (NMT), the quality of machine transla- tion has significantly increased. Its main advantage is that it improves translation efficiency over previ- ous machine translation techniques. However, the interpretability of NMT is poor: errors are difficult to interpret, i.e., to trace back to the training data. Past research (Federico et al., 2012) already pro- posed to combine the advantages of TM (inter- pretability) with MT (efficiency). Recently, meth- ods have been proposed to achieve closer integra- tion with NMT. For example, an additional encoder can be added to an NMT architecture specifically for TM matches (Cao and Xiong, 2018). The decoding algorithm can be modified to incorporate retrieved strings (Gu et al., 2018). An easy-to-implement TM–NMT integration has been proposed by (Bult´e and Tezcan, 2019): they concatenate the target- language side of matches retrieved from a TM with the sentence to translate. This only involves data pre-processing and augmentation. This is also com- patible with different NMT architectures. All of the approaches above were shown to lead to a significant increase in the quality of MT outputs. 2 Method In this paper, we propose to make use of the TM principle in conjunction with an NMT system, with- out altering the model architecture of the NMT sys- tem. In this way, the method can apply to any neu- ral network architecture and can leverage pre-trained models. Figure 1 illustrates this method. Suppose that we want to translate a sentence from English to German. We firstly retrieve English sen- tences from the parallel aligned data and obtain some similar English sentences which cover the in- put English sentence. For instance, …

Content truncated to 500 words