Word Rewarding for Adequate Neural Machine Translation

Word Rewarding for Adequate Neural Machine Translation Yuto Takebayashi†, Chu Chenhui‡, Yuki Arase†, Masaaki Nagata∗ †Graduate School of Information Science and Technology, Osaka University ‡Institute for Datability Science, Osaka University ∗NTT Communication Science Laboratories, NTT Corporation {takebayashi.yuto, arase}@ist.osaka-u.ac.jp, chu@ids.osaka-u.ac.jp, nagata.masaaki@lab.ntt.co.jp Abstract To improve the translation adequacy in neural machine translation (NMT), we propose a rewarding model with tar- get word prediction using bilingual dictionaries inspired by the success of decoder constraints in statistical machine translation. In particular, the model first predicts a set of tar- get words promising for translation; then boosts the proba- bilities of the predicted words to give them better chances to be output. Our rewarding model minimally interacts with the decoder so that it can be easily applied to the de- coder of an existing NMT system. Extensive evaluation under both resource-rich and resource-poor settings shows that (1) BLEU score improves more than 10 points with or- acle prediction, (2) BLEU score improves about 1.0 point with target word prediction using bilingual dictionaries cre- ated either manually or automatically, (3) hyper-parameters of our model are relatively easy to optimize, and (4) under- generation problem can be alleviated in exchange for increas- ing over-generated words. 1. Introduction Neural machine translation (NMT) [1, 2, 3] has dramatically improved machine translation quality compared to statistical machine translation (SMT). However, current NMT systems still suffer from the adequacy problem due to inappropriate lexical choice, under-generation, and over-generation [4]. In SMT, bilingual dictionaries have been used to improve ade- quacy in translation as decoder constraints. Typical example is the XML markup function implemented on MOSES [5]. Inspired by the decoding constraints for SMT, we pro- pose a rewarding model using bilingual dictionaries to ad- dress the adequacy problem in NMT. Our model rewards tar- get words that are promising to be used in correct translations by boosting their probabilities to be output by a decoder. It predicts such target words using bilingual dictionaries that are created manually or automatically. By applying byte pair encoding (BPE) [6] to dictionaries, our model can benefit from both BPE and dictionaries. While previous studies incorporate bilingual dictionaries into NMT for translation of rare words [7, 8] and domain- specific words [9], we do so to improve the adequacy of NMT. Hence, dictionaries are made use of translating not only specific types of words but also all words. In addition, these are methodologically different; our model simply bi- ases the trained decoder while previous models change the inside NMT architectures and require training of the entire systems. Due to this design, our model is easy to add to trained NMT systems and compatible with BPE. Extensive evaluation on Japanese-to-English and English-to-Japanese translation has been conducted using two datasets; IWSLT (TED Talk) [10], spoken language domain with a small set of bilingual sentences (223k), and ASPEC [11], a scientific domain with a large set of bilingual sentences (3M). We refer to the former as a resource-poor domain and the latter as a resource-rich domain, hereafter. The results …

Content truncated to 500 words