sylviawangfr / treatkgc

SICKLE

1. Data

• We use three datasets: DBped-P, DB15K, and NELL.
• Unzip resources/NELL.zip to resources/NELL/
• Unzip resources/DB15K.zip to resources/DB15K/

Data format:

• Triples: filename abox_hrt_uri.txt, h, r and t, seperated by tab
• Ontology: ontology in OWL2, n-triples format, containing at lease class hierarchy, disjoint class, domain/range etc.
• The L-method need literals, please refer to resources/NELL for data formats.
• The ACC need entity2type.txt, please refer to resources/NELL for data formats.

2. Dependencies

• Install pytorch according to your cpu/gpu environment.
• Install all packages in requirements.txt. Feel free to use conda, pyenv or others.
• Install java, jdk, maven

3. Preparing Konclude reasoner

mkdir java_owlapi/Konclude
unzip resources/packages/Konclude-v0.7.0-1135-Linux-x64-GCC-Static-Qt5.12.10.zip -d java_owlapi
mv java_owlapi/Konclude-v0.7.0-1135-Linux-x64-GCC-Static-Qt5.12.10 java_owlapi/Konclude 

4. Preparing TBox scanner

cd java_owlapi/TBoxTREAT
mvn clean install

5. Triple producer

source setup.sh
cd pipeline
python experiments.py --dataset=DB15K --work_dir="../outputs/proDB15K/"  --produce=True --start_acc=True --silver_eval=True --pred_type=False --pipeline=a_m_l --loops=1 --rel_model=complex --inductive=False --parallel=True --schema_aware_sampling=False 

"a_m_l" means run AnyBURL, materialization, blp together. If you only run single method, just use option a, m or l. --inductive=True would set blp to literal embedding. --inductive=False would set blp to pure KG embedding. Please refer to blp paper for more information.

For type prediction:

• Set --pred_type=True. This will run type prediction after the link prediction step.
• Or run

cd module_utils
python type_producer.py --dataset=DB15K --work_dir="../outputs/proDB15K/a_m_l/

6. For custom datasets:

• Please refer to the required data format.
• You need to generate TBox inconsistency justification patterns before you run the pipeline:

scripts/tbox_Scanner.sh  schema_file work_dir

---

7. TREAT downstream sampling (For my co-worker's project. You don't need this)

source setup.sh
cd ../../pipeline 
python TREAT_downstream.py

There are a few output files in outputs/treat_downstream/: \

• blp_new_triples.csv: all produced triples with scores.
• sample_and_score.pt: train sampling file, with format h,r,t,s.
• valid_hrt.txt: triples that consistent with schema.
• invalid_hrt.txt: triples that consistent with schema. The entities and relations in these files are indexed by abox_scanner/abox_utils.py. However, it can be translated to URI with minor changes in a pipeline.