The Model QLogicE

This repository is the implementation codes for the model QLogicE. This model joints the translation embedding TransE and the quantum logic E2R for the sake of capturing more features to improve the expressiveness of the knowledge graph completion task.

Download the Code

Firstly, the code can be cloned from the github repository by following command:

git clone https://github.com/gzupanda/QLogicE.git

The code mainly depends on PyTorch 1.1.0 or later versions and Python 3 which is modified from the baseline model E2R. In this code, there are seven directories corresponding to each dataset.

Datasets for Evalating the Model

The datasets used in our experiments are as follows.

Dataset	Entity	Relation	Train	Valid	Test	Triples
Kinship	104	26	8,544	1,068	1,074	10,686
UMLS	135	49	5,216	652	661	6,529
FB15k	14,951	1,345	483,142	50,000	59,071	592,213
WN18	40,943	18	141,442	5,000	5,000	151,442
FB15k237	14,505	237	272,115	17,535	20,466	310,116
WN18RR	40,943	11	86,835	3,034	3,134	93,003
YAGO3-10	123,182	37	1,079,040	5,000	5,000	1,089,040

In other words, every directory is an independent one with the data such as training, validing testing and trained model data. Besides, there are codes including data parser, training, testing, evaluating, etc. To better understand how to quick start to run the experiments, we take the dataset UMLS for example.

Run the Experiments

There are seven directories in this repository corresponding the seven datasets. They are devided as three groups, . Fortunately, the running is simple. For every dataset, we just run the following steps and take the dataset umls for example:

1. Enter the Directory

In every directory, there are eight python files and two directories. We can go into the objective directory to run the right codes. In this case, it is the directory unls.

cd QLogicE
cd umls

2. Training the Model

When we finish the above step, we can run the following command to training the dataset. And the trained model is saved in the directory save. In this case, it is the dataset UMLS.

python Train.py

3. Testing the Model

In this step, the code is used for test the training results. We only need to run the followinng command.

python Test.py

Model	FB15k	FB15k	FB15k	WN18	WN18	WN18	FB15k237	FB15k237	FB15k237	WN18RR	WN18RR	WN18RR	YAGO3-10	YAGO3-10	YAGO3-10
MRR	Hits@1	Hits@10	MRR	Hits@1	Hits@10	MRR	Hits@1	Hits@10	MRR	Hits@1	Hits@10	MRR	Hits@1	Hits@10
TransE	0.628	49.36	84.73	0.646	40.56	94.87	0.310	21.72	49.65	0.206	2.79	49.52	0.501	40.57	67.39
E2R	0.964	96.40	96.40	0.710	71.10	71.10	0.584	58.40	58.40	0.477	47.70	47.70	0.830	83.00	83.00
QLogicE	0.969	96.93	96.93	0.927	92.70	92.70	0.949	94.89	94.89	0.928	92.79	92.79	0.937	93.74	93.74

Performance of the Results

In this model, the proposed model achieves outstanding results, especially on the challenging datasets such as FB15k237, WN18RR and YAGO3-10.

Model	FB15k	FB15k	FB15k	WN18	WN18	WN18	FB15k237	FB15k237	FB15k237	WN18RR	WN18RR	WN18RR	YAGO3-10	YAGO3-10	YAGO3-10
MRR	Hits@1	Hits@10	MRR	Hits@1	Hits@10	MRR	Hits@1	Hits@10	MRR	Hits@1	Hits@10	MRR	Hits@1	Hits@10
ConvE	0.688	59.46	84.94	0.945	93.89	95.68	0.305	21.90	47.62	0.427	38.99	50.75	0.488	39.93	65.75
ConvKB	0.211	11.44	40.83	0.709	52.89	94.89	0.230	13.98	41.46	0.249	5.63	52.50	0.420	32.16	60.47
ConvR	0.773	70.57	88.55	0.950	94.56$	95.85	0.346	25.56	52.63	0.467	43.73	52.68	0.527	44.62	67.33
CapsE	0.087	1.934	21.78	0.890	84.55	95.08	0.160	7.34	35.60	0.415	33.69	55.98	0.000	0.00	0.00
RSN	0.777	72.34	87.01	0.928	91.23	95.10	0.280	19.84	44.44	0.395	34.59	48.34	0.511	42.65	66.43
TransE	0.628	49.36	84.73	0.646	40.56	94.87	0.310	21.72	49.65	0.206	2.79	49.52	0.501	40.57	67.39
STransE	0.543	39.77	79.60	0.656	43.12	93.45	0.315	22.48	49.56	0.226	10.13	42.21	0.049	3.28	7.35
CrossE	0.702	60.08	86.23	0.834	73.28	95.03	0.298	21.21	47.05	0.405	38.07	44.99	0.446	33.09	65.45
TorusE	0.746	68.85	83.98	0.947	94.33	95.44	0.281	19.62	44.71	0.463	42.68	53.35	0.342	27.43	47.44
RotatE	0.791	73.93	88.10	0.949	94.43	96.02	0.336	23.83	53.06	0.475	42.60	57.35	0.498	40.52	67.07
DistMult	0.784	73.68	86.32	0.824	72.60	94.61	0.313	22.44	49.01	0.433	39.68	50.22	0.501	41.26	66.12
ComplEx	0.848	81.56	90.53	0.949	94.53	95.50	0.349	25.72	52.97	0.458	42.55	52.12	0.576	50.48	70.35
Analogy	0.726	65.59	83.74	0.934	92.61	94.42	0.202	12.59	35.38	0.366	35.82	38.00	0.283	19.21	45.65
SimplE	0.726	66.13	83.63	0.938	93.25	94.58	0.179	10.03	34.35	0.398	38.27	42.65	0.453	35.76	63.16
HolE	0.800	75.85	86.78	0.938	93.11	94.94	0.303	21.37	47.64	0.432	40.28	48.79	0.502	41.84	65.19
TuckER	0.788	72.89	88.88	0.951	94.64	95.80	0.352	25.90	53.61	0.459	42.95	51.40	0.544	46.56	68.09
QLogicE	0.969	96.93	96.93	0.914	91.42	91.42	0.949	94.89	94.89	0.928	92.79	92.79	0.937	93.74	93.74

From this table, we can see that the proposed model QLogicE achieves the state-of-the-art over the existing models. Notably, the performance on dateset WN18 is not as promsing as the models achieved, though the model achieve the outstanding performance on other datasets such as FB15k, FB15k237 and WN18RR. Fortunately, these performance results are better than the baseline model E2R with large margin on dataset WN18. The ablation performance results show this point.

From this table, it can be seen that the proposed model significant better than the baselines model on datasets except WN18.

The Dense Feature Model

Inspired by the surprising performance of the model QLogicE, we coformulate a dense feature model framework from the perspective of information theory to guide the further knowledge graph completion model design. The key conception of it is expressive density, it is the criterion of dense feature model and we compute out the critical value for this end. It also show that how much room to be improved for a model on a specific dataset.

Data Related

the file triples.tex re

In the directories of datasets FB15k237, WN18RR and YAGO3-10, there are text file named triples in corresponding directories. This file including the all triples from training set, validing set and testing set. This is because our model runs under the assumption of close world, the entity and relation out of knowledge graph are incapable of link predicting.

the .zip file

Since the limitation of github, we have to compress the files more than 25MB into .zip file to upload them to the site. As a result, when run these model, you have to uncompress them.

License

This software comes under a non-commercial use license, please see the LICENSE file.