The model RosE

This repository contains the code of the main experiments presented in the paper:RosE: Towards the Expressiveness of Translation-based Model for Knowledge Graph Completion.

In this paper, we introduce freedoms to improve the expressiveness of the translation-based model by rotating and scaling the translation equation (i.e., the approximate equation $\mathbf{h} + \mathbf{r}\approx \mathbf{t}$). Further research shows that the manner of the tranformations of it strongly related with the expressiveness of the model.

Install

Firstly, the code of the model $\textit{RosE}$ can be cloned from the repository in github:

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

Secondly, the code depends on downhill, a theano-based Stochastic Gradient Descent implementation. The proposed model and baseline models are implemented in Python and the deep learning framework Theano on the basis of the model ComplEx.

To run the code, the dependencies should also be installed:

pip install -r requirements.txt

Furthermore, the code is compatible with both Python 2 and 3.

Run the experimental code

We provide two versions of the proposed model, they are RosE${1l}$ and RosE${2l}$, respectively.

1. on datasets Kinships and UMLS

These tow datasets are the small datasets and used for evaluating the model in the early age. For the dataset Kinships:

python kinships_1L.py
python kinships_2L.py

And for the dataset UMLS:

python umls_1L.py
python umls_2L.py

2. on datasets FB15k and WN18

These two datasets can be considered as standard ones for new embedding-based knowledge graph completion model evaluting. For the dataset Freebase (FB15k):

python fb15k_1L.py
python fb15k_2L.py

And for the dataset Wordnet (WN18):

python wn18_1L.py
python wn18_2L.py

In the following table, the results can be contributed into two groups. The first group includes the models from TransE to GTrans-GD, they are the models transformed from the tranlation equation linearly. The second group contain the model RotatE, RatE and TimE, which are newly proposed models transformed from the translation equation non-linearly.

Model	Stra.	FB15k	FB15k	FB15k	FB15k	WN18	WN18	WN18	WN18
TransE	unif.	125	243	47.1	34.9	251	263	89.2	75.4
TransH	unif.	84	211	58.5	42.5	303	318	86.7	75.4
TransH	bern.	87	212	64.4	45.7	388	401	82.3	73.0
TransR	unif.	78	226	65.5	43.8	219	232	91.7	78.3
TransR	bern.	77	198	68.7	48.2	225	238	92.0	79.8
CTransR	unif.	82	233	66.3	44.0	230	243	92.3	78.9
CTransR	bern.	75	199	70.2	48.4	218	231	92.3	79.4
TransD	unif.	67	211	74.2	49.4	229	242	92.5	79.2
TransD	bern.	91	194	77.3	53.4	212	224	92.2	79.6
TransPES	unif.	66	198	67.3	48.1	212	223	81.3	71.6
TransE-RS	unif.	62	$\mathbf{161}$	72.3	53.1	348	362	93.7	80.3
TransE-RS	bern.	63	$\mathbf{161}$	72.1	53.2	371	385	93.7	80.4
TransH-RS	unif.	64	$\underline{163}$	72.6	53.4	389	401	94.7	81.2
TransH-RS	bern.	77	178	75.0	53.6	357	371	94.5	80.3
TransGH	unif.	66	186	79.8	$\underline{54.0}$	$\underline{179}$	$\underline{191}$	94.8	$\underline{81.4}$
TransGH	bern.	64	186	80.1	$\mathbf{54.1}$	197	210	$\underline{95.3}$
GTrans-DW	unif.	147	256	63.4	44.1	197	210	92.2	78.4
GTrans-DW	bern.	126	235	60.5	43.1	$\mathbf{166}$	$\mathbf{180}$	90.3	77.1
GTrans-SD	unif.	66	207	75.1	50.6	234	247	92.9	79.1
GTrans-SD	bern.	85	189	75.3	52.9	202	215	93.5	80.2
RotatE	unif.	$\underline{40}$	-	74.6	-	309	-	95.9	-
RatE	unif.	$\mathbf{24}$	-	$\mathbf{89.8}$	-	180	-	$\mathbf{96.2}$	-
TimE	unif.	45.9	-	$\underline{87.9}$	-	259	-	$\underline{96.1}$	-
RosE$_{1l}$	unif.	79	233	81.5	50.4	271	284	95.0	$\underline{81.4}$
RosE$_{2l}$	unif.	79	234	80.4	48.7	347	362	94.5	78.8

on the datasets FB15k237 and WN18RR

These two datasets consider to be challenging since they are proposed for test set leakage. For the dataset Freebase (FB15k237):

python fb15k237_1L.py
python fb15k237_2L.py

And for the dataset Wordnet (WN18RR):

python wn18rr_1L.py
python wn18rr_2L.py

To run on GPU (approx 5x faster at least), simply add the following theano flag before the python call:

THEANO_FLAGS='device=gpu' python fb15k_1L.py

Run on your own data

Create a subfolder in the datasets folder, and put your data in three files train.txt, valid.txt and test.txt. Each line is a triple, in the format: 

subject_entity_id	relation_id	object_entity_id

separated with tabs. Then modify fb15k_1L.py (fb15k_2L.py) for example, by changing the name argument in the build_data function call to your data set folder name:

fb15kexp = build_data(name = 'your_dataset_folder_name',path = tools.cur_path + '/datasets/')

License

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