Abstract
Neural Machine Translation (NMT) can be used to generate fluent output. Assuch, language models have been investigated for incorporation with NMT. Inprior investigations, two models have been used: a translation model and alanguage model. The translation model's predictions are weighted by thelanguage model with a hand-crafted ratio in advance. However, these approachesfail to adopt the language model weighting with regard to the translationhistory. In another line of approach, language model prediction is incorporatedinto the translation model by jointly considering source and targetinformation. However, this line of approach is limited because it largelyignores the adequacy of the translation output. Accordingly, this work employs two mechanisms, the translation model and thelanguage model, with an attentive architecture to the language model as anauxiliary element of the translation model. Compared with previous work inEnglish--Japanese machine translation using a language model, the experimentalresults obtained with the proposed Dynamic Fusion mechanism improve BLEU andRank-based Intuitive Bilingual Evaluation Scores (RIBES) scores. Additionally,in the analyses of the attention and predictivity of the language model, theDynamic Fusion mechanism allows predictive language modeling that conforms tothe appropriate grammatical structure.