search

nju-websoft / JAPE

Cross-Lingual Entity Alignment via Joint Attribute-Preserving Embedding, ISWC 2017
http://ws.nju.edu.cn/jape/
MIT License
98 stars 22 forks source link

The problem of training pair set (sup_ent_ids). #9

Open pp413 opened 3 years ago

pp413 commented 3 years ago

Did no one find that the file "sup_ent_ids" was wrong? The two entity ids in each entity pair of sup_ent_ids are same.

sunzequn commented 3 years ago

Hi,

Thanks for your interest in our work.

This is not a mistake. The file "sup_ent_ids" contains IDs for seed entity alignment as training data. Each pair of aligned entities uses the same ID to merge their representations. In this way, two aligned entities are actually one identify in the model such that we do not need an additional alignment learning loss to minimize their distance.