Knowledge Graph Enhanced Neural Machine Translation via Multi-task Learning on Sub-entity Granularity

Proceedings of the 28th International Conference on Computational Linguistics, pages 4495–4505 Barcelona, Spain (Online), December 8-13, 2020 4495 Knowledge Graph Enhanced Neural Machine Translation via Multi-task Learning on Sub-entity Granularity Yang Zhao1,2, Lu Xiang1,2, Junnan Zhu1,2, Jiajun Zhang1,2, Yu Zhou1,2,3, and Chengqing Zong1,2, 1National Laboratory of Pattern Recognition, Institute of Automation, CAS 2University of Chinese Academy of Sciences 3Beijing Fanyu Technology Co., Ltd {yang.zhao,lu.xiang,junnan.zhu,jjzhang}@nlpr.ia.ac.cn {yzhou,cqzong}@nlpr.ia.ac.cn Abstract Previous studies combining knowledge graph (KG) with neural machine translation (NMT) have two problems: i) Knowledge under-utilization: they only focus on the entities that appear in both KG and training sentence pairs, making much knowledge in KG unable to be fully utilized. ii) Granularity mismatch: the current KG methods utilize the entity as the basic granularity, while NMT utilizes the sub-word as the granularity, making the KG different to be utilized in NMT. To alleviate above problems, we propose a multi-task learning method on sub-entity granularity. Specifically, we first split the entities in KG and sentence pairs into sub-entity granularity by using joint BPE. Then we utilize the multi-task learning to combine the machine translation task and knowledge reasoning task. The extensive experiments on various translation tasks have demonstrated that our method significantly outperforms the baseline models in both translation quality and handling the entities. 1 Introduction Neural machine translation (NMT) based on the encoder-decoder architecture becomes a new state-of- the-art approach due to its distributed representation and end-to-end learning (Luong et al., 2015; Gehring et al., 2017; Vaswani et al., 2017). During translation, the translation quality of the entities in a sentence has a great influence on the translation quality of the whole sentence. However, translating these entities remains challenging (Mous- sallem et al., 2019) and various methods are proposed to improve the translation of entities (Zhang and Zong, 2016; Dinu et al., 2019; Ugawa et al., 2018; Wang et al., 2019). Among them, some methods aim at incorporating the knowledge graph (KG) to utilize their structured knowledge on entities and improve the entity translation. These studies utilize KG to enhance the semantic representing of entities in a sen- tence (Moussallem et al., 2019; Lu et al., 2018) or extract the important semantic vectors with KG (Shi et al., 2016). Although great efforts have been made to incorporate KG into NMT, we find the existing methods have the following two problems: Knowledge Under-utilization: Given a KG (denoted by K) and a parallel sentence pair dataset (denoted by D), the full entity set U can be divided into four subsets as shown in Fig. 1 (a): 1) K ∩ D entities, which appear in both K and D; 2) D−K entities, which only appear in D; 3) K −D entities, which only appear in K; 4) U −(K∪D) entities, which are neither in K nor D. While previous studies (Shi et al., 2016; Moussallem et al., 2019; Lu et al., 2018) only focus on the K ∩ D entities, the other three subsets are ignored. Consequently, much knowledge on the other three subsets in …

Content truncated to 500 words