Neural Machine Translation Integrated with Knowledge Graph

2021 Fiscal Year Final Research Report

• Project/Area Number: 20K23325
• Research Category: Grant-in-Aid for Research Activity Start-up
• Allocation Type: Multi-year Fund
• Review Section: 1001:Information science, computer engineering, and related fields
• Research Institution: Nara Institute of Science and Technology
• Principal Investigator: Watanabe Taro 奈良先端科学技術大学院大学, 先端科学技術研究科, 教授 (90395038)
• Project Period (FY): 2020-09-11 – 2022-03-31
• Keywords: 機械翻訳 / 知識グラフ

Outline of Final Research Achievements

Neural machine translation demands huge data when training the translation model, although its performance has been drastically improved by deep learning. However, simply increasing training data does not assure that the trained model can fluently translate named entities, properties, e.g., date of birth, or relations with other objects, e.g., affiliations, since such knowledge will be updated almost every day. This work investigates a method to solve the issue by integrating multilingual a knowledge graph into machine translation, which is knowledge representation denoting attributes and relations of objects with partially multilingual annotation.
In this research, we proposed a machine translation model which integrates representations from knowledge graph that is trained by subword unit, not word-wise unit. Experimental results on machine translation tasks showed that named entities are translated correctly after our manual investigations.

Free Research Field

自然言語処理

Academic Significance and Societal Importance of the Research Achievements

本研究における知識グラフと機械翻訳を統合したモデルにより、知識を反映した機械翻訳を実現した。本手法により、人名や地名等の固有表現をより正しく翻訳できることを示している。今後は、知識グラフを更新することで機械翻訳モデルの再学習を全く必要としない機械翻訳モデルを実現することにより、各ドメインへと容易に適用可能、かつ、カスタマイズ可能なシステムの実現を目指す。