[3Xin4] Poster session 1

[3Xin4-27] A Comparative Study on the Use of Named Entity in Neural Machine Translation

〇Naoki Minamibata¹, Akihiro Tamura¹, Tsuneo Kato¹ (1.Doshisha University)

Keywords:Neural Machine Translation, Named Entity, Transformer

The performance of neural machine translation (NMT) has been improved by utilizing named entity (NE) in source and target language sentences. Two promising methods for NE-based NMT have been proposed: (i) “a tagging method” that inserts NE tags into source and target language sentences and (ii) “an embedding method” that adds NE embeddings to word embeddings in the embedding layers of an encoder and a decoder. To date, these two methods have not been compared. This study experimentally compares the two methods on the WMT2014 English-to-German, WMT2014 German-to-English WMT2020 English-to-Japanese, and WMT2020 Japanese-to-English translation tasks. The experimental results showed that the tagging method is better at translating sentences containing NEs while the embedding method is better at translating sentences not containing NEs. Through the experiments, it was found that overall translation performance can be improved by switching between the two methods depending on whether an input sentence contains NEs or not.