New: Add Evaluation code for DKRL(CNN)+TransE, additional TransE results are needed to run this evaluation. "../transE_res/entity2vec."+version "../transE_res/relation2vec."+version with the same dimension and unif/bern.
Description-Embodied Knowledge Representation Learning (DKRL)
Representation Learning of Knowledge Graphs with Entity Descriptions (AAAI'16)
Ruobing Xie
Just type make in the folder ./
Pre-trained embeddings for entity/relation/word are optional. We update both Structure-based Representations and Description-based Representations in this version. We can also fix Structure-based Representations pre-trained by other models and only update Description-based Representations.
For DKRL, we learn structure-based representations (SBR) and description-based representations (DBR) simultaneously in training. However, Test_cnn.cpp only use description-based representations for prediction. You can load in both entity representations for joint prediction.
FB15k is published by the author of the paper "Translating Embeddings for Modeling Multi-relational Data (2013)." [download] You can also get FB15k from here: [download]
Entity list and descriptions of FB15k used in this work [download]
FB20k is based on FB15k and used for zero-shot scenario [download]
Entity type information for entity classification [download]
All these datasets are also in data.rar.
Entity name file [download]
If the codes or datasets help you, please cite the following paper:
Ruobing Xie, Zhiyuan Liu, Jia Jia, Huanbo Luan, Maosong Sun. Representation Learning of Knowledge Graphs with Entity Descriptions. The 30th AAAI Conference on Artificial Intelligence (AAAI'16).