RETA

RETA (as it suggests the Relation and Tail for a given head entity) is an end-to-end solution tackling instance completion problem over KGs. It consists of two components: a RETA-Filter and RETA-Grader. More precisely, our RETA-Filter first generates candidate relation-tail pairs for a given head by extracting and leveraging the schema of a KG; our RETA-Grader then evaluates and ranks the candidate relation-tail pairs considering the plausibility of both the candidate triplet and its corresponding schema using a newly-designed KG embedding model. ​

How to run the code

Datasets link: http://bit.ly/3t2WFTE

Train and Evaluate JF17k (use the pre-processed dataset: http://bit.ly/3tksFCK)

python main.py --indir data/JF17k --withTypes True --epochs 3000 --batchsize 128 --num_filters 50 --embsize 100 --learningrate 0.0002 --outdir data/JF17k --gpu_ids 0 --num_negative_samples 1 --buildTypeDictionaries True --sparsifier 1

python main.py --indir data/JF17k --withTypes True --epochs 3000 --batchsize 128 --num_filters 50 --embsize 100 --learningrate 0.0002 --outdir data/JF17k/RETA_batchSize128_epoch3000_embSize100_lr0.0002_sparsifier1_numFilters50 --load True --gpu_ids 3 --num_negative_samples 1 --atLeast 2 --topNfilters -10 --buildTypeDictionaries True --sparsifier 1 --entitiesEvaluated both

python main.py --indir data/JF17k --withTypes True --epochs 3000 --batchsize 128 --num_filters 50 --embsize 100 --learningrate 0.0002 --outdir data/JF17k/RETA_batchSize128_epoch3000_embSize100_lr0.0002_sparsifier1_numFilters50 --load True --gpu_ids 3 --num_negative_samples 1 --atLeast 1 --topNfilters -10 --buildTypeDictionaries True --sparsifier 1 --entitiesEvaluated one

python main.py --indir data/JF17k --withTypes True --epochs 3000 --batchsize 128 --num_filters 50 --embsize 100 --learningrate 0.0002 --outdir data/JF17k/RETA_batchSize128_epoch3000_embSize100_lr0.0002_sparsifier1_numFilters50 --load True --gpu_ids 3 --num_negative_samples 1 --atLeast 2 --topNfilters -10 --buildTypeDictionaries True --sparsifier 1 --entitiesEvaluated none

Train and Evaluate FB15k (use the pre-processed dataset: http://bit.ly/3cHacuq)

python main.py --indir data/FB15k --withTypes True --epochs 200 --batchsize 128 --num_filters 200 --embsize 100 --learningrate 0.0002 --outdir data/FB15k --gpu_ids 7 --num_negative_samples 1 --buildTypeDictionaries True --sparsifier 1

python main.py --indir data/FB15k --withTypes True --epochs 200 --batchsize 128 --num_filters 200 --embsize 100 --learningrate 0.0002 --outdir data/FB15k/RETA_batchSize128_epoch200_embSize100_lr0.0002_sparsifier1_numFilters200 --load True --gpu_ids 3 --num_negative_samples 1 --atLeast 2 --topNfilters -10 --buildTypeDictionaries True --sparsifier 1 --entitiesEvaluated both

python main.py --indir data/FB15k --withTypes True --epochs 200 --batchsize 128 --num_filters 200 --embsize 100 --learningrate 0.0002 --outdir data/FB15k/RETA_batchSize128_epoch200_embSize100_lr0.0002_sparsifier1_numFilters200 --load True --gpu_ids 3 --num_negative_samples 1 --atLeast 1 --topNfilters -10 --buildTypeDictionaries True --sparsifier 1 --entitiesEvaluated one

python main.py --indir data/FB15k --withTypes True --epochs 200 --batchsize 128 --num_filters 200 --embsize 100 --learningrate 0.0002 --outdir data/FB15k/RETA_batchSize128_epoch200_embSize100_lr0.0002_sparsifier1_numFilters200 --load True --gpu_ids 3 --num_negative_samples 1 --atLeast 2 --topNfilters -10 --buildTypeDictionaries True --sparsifier 1 --entitiesEvaluated none

Train and Evaluate humans_wikidata (use the pre-processed dataset: http://bit.ly/2M1S1ER)

python main.py --indir data/humans_wikidata --withTypes True --epochs 3000 --batchsize 128 --num_filters 100 --embsize 100 --learningrate 0.0002 --outdir data/humans_wikidata --gpu_ids 2 --num_negative_samples 1 --buildTypeDictionaries True --sparsifier 4

python main.py --indir data/humans_wikidata --withTypes True --epochs 3000 --batchsize 128 --num_filters 100 --embsize 100 --learningrate 0.0002 --outdir data/humans_wikidata/RETA_batchSize128_epoch3000_embSize100_lr0.0002_sparsifier4_numFilters100 --load True --gpu_ids 3 --num_negative_samples 1 --atLeast 1 --topNfilters -80 --buildTypeDictionaries True --sparsifier 4 --entitiesEvaluated both

python main.py --indir data/humans_wikidata --withTypes True --epochs 3000 --batchsize 128 --num_filters 100 --embsize 100 --learningrate 0.0002 --outdir data/humans_wikidata/RETA_batchSize128_epoch3000_embSize100_lr0.0002_sparsifier4_numFilters100 --load True --gpu_ids 3 --num_negative_samples 1 --atLeast 1 --topNfilters -80 --buildTypeDictionaries True --sparsifier 4 --entitiesEvaluated one

Parameter setting:

In main.py, you can set:

--indir: input file directory

--withTypes: True trains RETA, False trains RETA no type

--epochs: number of training epochs

--batchsize: batch size of training set

--num_filters: number of filters used in the CNN

--embsize: embedding size

--load: load a pre-trained RETA model and evaluate

--learningrate: learning rate

--outdir: where to store RETA model

--num_negative_samples: number of negative samples

--gpu_ids: gpu to be used for train and test the model

--atLeast: beta parameter

--topNfilters: alpha parameter

--buildTypeDictionaries: store dictionaries to speed-up the code

--sparsifier: top-k types for each entity

--entitiesEvaluated: both (evaluate facts where both h and t have types), one (evaluate facts where either h or t have types), none (evaluate facts where both h and t don't have types). In our paper we have computed the weighted average between these three settings.

Data preprocessing example

python build_dataset.py --indir FB15k --atLeast 1000
python builddata.py --data_dir data/FB15k/

Reference

If you use our code or datasets, please cite:

Rosso, P., Yang, D., Ostapuk, N., Cudré-Mauroux, P. (2021, April). RETA: A Schema-Aware, End-to-End Solution for Instance Completion in Knowledge Graphs. In Proceedings of The Web Conference 2021.