Knowledge Graph Completion

This repository holds the code and datasets for experiments in the paper (Please, cite these publication if you use our source codes and datasets) Cite:

@inproceedings{zhao-etal-2020-connecting,
    title = "Connecting Embeddings for Knowledge Graph Entity Typing",
    author = "Zhao, Yu  and
      zhang, anxiang  and
      Xie, Ruobing  and
      Liu, Kang  and
      WANG, Xiaojie",
    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
    month = jul,
    year = "2020",
    address = "Online",
    publisher = "Association for Computational Linguistics",
    url = "https://www.aclweb.org/anthology/2020.acl-main.572",
    pages = "6419--6428",
    abstract = "Knowledge graph (KG) entity typing aims at inferring possible missing entity type instances in KG, which is a very significant but still under-explored subtask of knowledge graph completion. In this paper, we propose a novel approach for KG entity typing which is trained by jointly utilizing local typing knowledge from existing entity type assertions and global triple knowledge in KGs. Specifically, we present two distinct knowledge-driven effective mechanisms of entity type inference. Accordingly, we build two novel embedding models to realize the mechanisms. Afterward, a joint model via connecting them is used to infer missing entity type instances, which favors inferences that agree with both entity type instances and triple knowledge in KGs. Experimental results on two real-world datasets (Freebase and YAGO) demonstrate the effectiveness of our proposed mechanisms and models for improving KG entity typing. The source code and data of this paper can be obtained from: https://github.com/Adam1679/ConnectE .",
}

There are three kinds of inference of missing data in a knowledge graph: 1) missing entity, 2) missing relation type and 3) missing entity type. Our code implement a E2T model and a E2T_TRT model to improve missing entity type prediction performance.

Author: Adam Zhang (anxiangz@andrew.cmu.edu)

Paper: https://www.aclweb.org/anthology/2020.acl-main.572/

Dependency Requirements: Python==3.7

The source codes are modified versions from original source code from

Usage instructions:

\1. Parameter Setup

• 1.1 Related files

    1.1.1 ./run_E2T_FB.py (for our E2T model on FB15k dataset)
  
    1.1.2 ./run_E2T_TRT_FB.py (for our E2T-TRT model on FB15k dataset)
  
    1.1.3 ./run_E2T_YAGO.py (for our E2T model on YAGO43k dataset)
  
    1.1.4 ./run_E2T_TRT_YAGO.py (for our E2T-TRT model on YAGO43k dataset)
  

  1.1.5 ./config/*/config.json

    All the files in the config directory are related to the parameter setup for the very model. For example, file "./config/FB15k/E2T/config.json" is for the parameter setup of the E2T model on the YAGO43k dataset.
  

• 1.2 Parameter setup

    All the files in the "config" directory contains the default parameter setup we use to train the model. For each parameter, the explanation is following:
  
    - nbatch: number of batches to train in one epoch. Generally, we use vectorization mechanism to speed up the training process. If you increase this parameter, then it means you train more batches in one round of training. Consequently, you have to simultaneously calculate a larger matrix in on batch.(For all our model, the default value is 20.)
    - learning_rate: the learning rate of the training model. (For all our model, the default value is 0.1.)
    - entity_dim: the dimension of the entity/relation vector.
    - type_dim: the dimension of the type vector.
    - norm: the kind of norm we use to implement the loss function. (For all our model, we use L2 norm.)
    - margin: Margin for loss function.
    - evaluator: the kind of evaluator to use when evaluating the model. (We only implement type prediction evaluator, which is "Type_Evaluator")
    - evaluator_time: number of epochs to train before one evaluation of the whole test sample. 
  

  \2. Inference of missing entity type (ETE model)

• 2.1 Related datasets

    2.1.1 Freebase: ../data/FB15k/origin/
  
    2.1.2 YAGO: ../datasets/YAGO/YAGO43k/
  
    2.1.3 ../datasets/YAGO/YAGO43k_Entity_Types/
  

• 2.2 How to run

  Before we run the model, we have to firstly clean the original dataset because the train/valid/test entity type data is constructed in a way that has some errors. Firstly, we should clean the entity type train data by erasing all the samples in the training set, in which the entity does not appear in the triplet training set. Second, we erasing all the samples in the valid/test set, in which the entity or entity type dose no appear in the entity type training set. Also, this would generate the type-relation-type dataset, which is needed for running E2T_TRT model.

    2.2.1 Prepare the dataset.
  
    python ./data/FB15k/prepare_data_FB.py
  
    python ./data/YAGO/prepare_data_YAGO.py
  

  2.2.2 Run the program

    python ./run_E2T_FB.py
  
    python ./run_E2T_TRT_FB.py
  
    python ./run_E2T_YAGO.py
  
    python ./run_E2T_TRT_YAGO.py
  

Note Sorry for the inconvenience. Due to some reasons, the author of this repo is too busy to clean the new version source code currently. He would update this repo once have time. Generally, the differences between the current version and the new version are about naming of some models some trivial modifications for some model and data preprocessing (It would not affect the performance too much according to the expiriments.)