Semantically Constitutive Entities in Knowledge Graphs

This repository contains the code for training Semantically Constitutive Entities (SemCon), as well as evaluation on the Link Prediction task.

Usage

We recommend creating a Conda environment named semcon via the provided environment.yml:

conda env create -f environment.yml
conda activate semcon

Learning Knowledge Graph Embeddings

We provide an example of using OpenKE for Knowledge Graph (KG) embedding training. Please refer to the CoDEx repository for instructions on obtaining the CoDEx dataset.

We first clone the OpenKE repository and insert TesterMaxRank, which assigns the maximum rank for all tied items:

git clone https://github.com/thunlp/OpenKE.git
cd OpenKE && git checkout 4b8adac
cp -ai ../learn_kg_embs/openke_mod/* openke/
cp openke/base/Base.cpp openke/base/Base.original
awk 'NR==6{print "#include \"TestMaxRank.h\""}1' openke/base/Base.original > openke/base/Base.cpp
cd openke && bash make.sh

The following commands trains (TransE/TransH/TransD) KG embeddings on the FB15K237 dataset using OpenKE-suggested hyperparameters, and saves them in the given directory (REPLACE_WITH_KG_DIR):

# TransE
cd learn_kg_embs
python learn_kg_embs.py --in_path ../OpenKE/benchmarks/FB15K237/ --out_path REPLACE_WITH_KG_DIR --nbatches 100 --threads 4 --sampling_mode normal --bern_flag 1 --filter_flag 1 --neg_ent 25 --neg_rel 0 --model TransE --dim 200 --p_norm 1 --norm_flag True --loss_margin 5 --train_times 1000 --alpha 1.0 --opt_method sgd 

# TransH
cd learn_kg_embs
python learn_kg_embs.py --in_path ../OpenKE/benchmarks/FB15K237/ --out_path REPLACE_WITH_KG_DIR --nbatches 100 --threads 4 --sampling_mode normal --bern_flag 1 --filter_flag 1 --neg_ent 25 --neg_rel 0 --model TransH --dim 200 --p_norm 1 --norm_flag True --loss_margin 4 --train_times 1000 --alpha 0.5 --opt_method sgd

# TransD
cd learn_kg_embs
python learn_kg_embs.py --in_path ../OpenKE/benchmarks/FB15K237/ --out_path REPLACE_WITH_KG_DIR --nbatches 100 --threads 4 --sampling_mode normal --bern_flag 1 --filter_flag 1 --neg_ent 25 --neg_rel 0 --model TransD --dim_e 200 --dim_r 200 --p_norm 1 --norm_flag True --loss_margin 4 --train_times 1000 --alpha 1.0 --opt_method sgd

Evaluating baseline performance

We first process the embeddings learnt in REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL} (e.g., FB15K237_TransE) , and save the output file {DATASET}_{KG_MODEL}-{MODEL}-{P}-{G}.cooX in the given directory (REPLACE_WITH_OUTPUT_DIR):

cd baselines
python -u train_kg_centrality_baselines.py --c_ij REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL} --output_dir REPLACE_WITH_OUTPUT_DIR --degree_type all

Next, we run the evaluation script:

cd eval
python kg_link_pred.py --c_ij REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL}/ --eval_X REPLACE_WITH_OUTPUT_DIR/{DATASET}_{KG_MODEL}-{MODEL}-{P}-{G}.cooX

Training Semantically Constitutive Entities

The following commands trains Semantically Constitutive Entities via a two-step procedure, and saves them in the given directory (REPLACE_WITH_DIR):

Note: Requires a Gurobi license to run.

cd semantically_constitutive
# compile programs
make
# Step 1: Create relaxed solution at REPLACE_WITH_DIR/RELAXED_COOX_FILE
./relaxed_partial -c REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL} -k 4000 -g 10 -o REPLACE_WITH_OUTPUT_DIR
# Step 2: Obtain integer solution, with REPLACE_WITH_DIR/RELAXED_COOX_FILE as initial solution
./relaxed_solver -c REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL} -x REPLACE_WITH_OUTPUT_DIR/COOX_FILE -k 4000 -g 10 -o REPLACE_WITH_OUTPUT_DIR

Evaluation on the Link Prediction task can be done as above:

cd eval
python kg_link_pred.py --c_ij REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL}/ --eval_X REPLACE_WITH_OUTPUT_DIR/COOX_FILE

---

Baselines

We provide example code for replicating the baseline models used in our paper. All output file are evaluated in the same manner as above.

KG Centrality Measures

cd baselines
python train_kg_centrality_baselines.py --c_ij REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL} --output_dir REPLACE_WITH_OUTPUT_DIR --degree_type group_in --p 4000 --g 10

SSA

We use the SSA implementation from @hungnt55.

cd baselines
git clone https://github.com/hungnt55/Stop-and-Stare.git
unzip Stop-and-Stare/SSA_release_2.1.zip
cp coox2bin.cpp SSA_release_2.0/SSA
cd SSA_release_2.0/SSA
echo -e "\tg++ coox2bin.cpp -o coox2bin \$(PARA)" >> Makefile
make
coox2bin REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL} REPLACE_WITH_OUTPUT_DIR/{DATASET}_{KG_MODEL}-{MODEL}.bin
./SSA -i REPLACE_WITH_OUTPUT_DIR/{DATASET}_{KG_MODEL}.bin -k 4000 -epsilon 0.03 -delta 0.01 -m LT > {DATASET}_{KG_MODEL}_SSA.out
python process_ssa_output.py --input {DATASET}_{KG_MODEL}_SSA.out --g 10 --output_dir REPLACE_WITH_OUTPUT_DIR

PageRank

We use the PageRank implementation from @louridas.

cd baselines
git clone https://github.com/louridas/pagerank.git
cd pagerank/cpp && make
./pagerank -n -a 0.85 -c 0.00000000000000000000000000000001 REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL} > REPLACE_WITH_OUTPUT_DIR/{DATASET}_{KG_MODEL}.pr
python process_pagerank_output.py --c_ij REPLACE_WITH_KG_DIR/{DATASET}_{KG_MODEL} --pr_out REPLACE_WITH_OUTPUT_DIR/{DATASET}_{KG_MODEL}.pr --p 4000 --g 10 --output_dir REPLACE_WITH_OUTPUT_DIR

Citiation

Our paper can be cited in the following formats:

APA

Chia, C., Tkachenko, M., & Lauw, H. (2022). Semantically Constitutive Entities in Knowledge Graphs. In Database and Expert Systems Applications: 34th International Conference, DEXA 2023.

Bibtex

@inproceedings{chia2023semantically,
    title={Semantically Constitutive Entities in Knowledge Graphs},
    author={Chia, Chong Cher and Tkachenko, Maksim and Lauw, Hady W},
    booktitle={Database and Expert Systems Applications: 34th International Conference, DEXA 2023, Penang, Malaysia, August 28-30, 2023, Proceedings},
    year={2023},
    publisher={Springer Nature}
}