Link Prediction Model for Automated Extraction of Hypernym and Hyponym Relations in Multiple Languages

In natural language processing, hypernym-hyponym relations are the core of the body of knowledge and are useful for many downstream tasks, for example, technical trend analysis and patent examination.However, it is very costly to manually maintain these relations between terms.In this paper, we extract hypernym-hyponym relations from patent text data written in Japanese, English, and Chinese, and automatically construct a multilingual thesaurus.The proposed method consists of the following two steps.First, we use a generative adversarial network (GAN) to identify the terms in a hypernym-hyponym relation.Then, ConvE and GraphSAGE are combined to predict links on the graph of hypernym-hyponym relations constructed in the previous step, and to predict missing edges that should be in a hypernym-hyponym relation.In experiments conducted to demonstrate the effectiveness of the proposed method, it was found that our method outperformed previous methods in both the identification of hypernymhyponym relations using GAN and link prediction using a combination of ConvE and GraphSAGE.We constructed a classifier that can discriminate between hypernym-hyponym relations using a trained Discriminator with GAN.In our experiments, we achieved a recall rate of 0.936 in English.And we propose a new method that can automatically complement missing hypernym-hyponym relations.The proposed model achieved a score of 97.3 on the H@10 evaluation index with respect to the prediction of Chinese hypernym-hyponym relations.