jRDF2Vec is a Java implementation of RDF2Vec.
It supports multi-threaded, in-memory (or disk-access-based) walk generation and training.
You can generate embeddings for any NT
, NQ
, OWL/XML
, RDF HDT
,
TDB 1
, or TTL
file.
Found a bug? Don't hesitate to open an issue.
How to cite?
Portisch, Jan; Hladik, Michael; Paulheim, Heiko. RDF2Vec Light - A Lightweight Approach for Knowledge Graph Embeddings. Proceedings of the ISWC 2020 Posters & Demonstrations. 2020. [to appear]
An open-access version of the paper is available here.
Download this project, execute mvn clean install
.
Alternatively, you can download the packaged JAR of the latest successful: commit
here.
- Java 8 or later.
- Python 3.8 or later with the dependencies described in requirements.txt installed.
(Conda users can directly use the environment.yml file.)
You can check if you set up the environment (Python 3 + dependencies) correctly by running:
java -jar jrdf2vec-1.1-SNAPSHOT.jar -checkInstallation
The command line output will list missing requirements or print Installation is ok ✔
.
Use the resulting jar from the target
directory.
Minimal Example
java -jar jrdf2vec-1.1-SNAPSHOT.jar -graph ./kg_file.hdt
-graph <graph_file>
The file containing the knowledge graph for which you want to generate embeddings. The<graph_file>
can be any triple file, HDT file, a directory which contains NT files, or a TDB1 directory.
jRDF2Vec follows the convention over configuration design paradigm to increase usability. You can overwrite the default values by setting one or more optional parameters.
Parameters for the Walk Configuration
-onlyWalks
If added to the call, this switch will deactivate the training part so that only walks are generated. If training parameters are specified, they are ignored. The walk generation also works with the-light
parameter.-light <entity_file>
If you intend to use RDF2VecLight, you have to use this switch followed by the file path ot the describing the entities for which you require an embedding space. The file should contain one entity (full URI) per line.-numberOfWalks <number>
(default:100
)
The number of walks to be performed per entity.-depth <depth>
(default:4
)
This parameter controls the depth of each walk. Depth is defined as the number of hops. Hence, you can also set an odd number. A depth of 1 leads to a sentence in the form<s p o>
.-walkGenerationMode <MID_WALKS | MID_WALKS_DUPLICATE_FREE | RANDOM_WALKS | RANDOM_WALKS_DUPLICATE_FREE>
(default for light:MID_WALKS
, default for classic:RANDOM_WALKS_DUPLICATE_FREE
)
This parameter determines the mode for the walk generation (multiple walk generation algorithms are available).-threads <number_of_threads>
(default:(# of available processors) / 2
)
This parameter allows you to set the number of threads that shall be used for the walk generation as well as for the training.-walkDirectory <directory where walk files shall be generated/reside>
The directory where the walks shall be generated into. In case of-onlyTraining
, the directory where the walks reside.-embedText
If added to the call, this switch will also generate walks that contain textual fragments of datatype properties.
Parameters for the Training Configuration
-onlyTraining
If added to the call, this switch will deactivate the walk generation part so that only the training is performed. The parameter-walkDirectory
must be set. If walk generation parameters are specified, they are ignored.-trainingMode <cbow | sg>
(default:sg
)
This parameter controls the mode to be used for the word2vec training. Allowed values arecbow
andsg
.-dimension <size_of_vector>
(default:200
)
This parameter allows you to control the size of the resulting vectors (e.g. 100 for 100-dimensional vectors).-minCount <number>
(default:1
)
This parameter controls the minimum word count for the word2vec training. Unlike in the gensim defaults, this parameter is set to 1 by default because for knowledge graph embeddings, a vector for each node/arc is desired.-noVectorTextFileGeneration
|-vectorTextFileGeneration
A switch which indicates whether a text file with the vectors shall be persisted on the disk. This is enabled by default. Use-noVectorTextFileGeneration
to disable the file generation.-sample <rate>
(default:0.0
)
The threshold for configuring which higher-frequency words are randomly downsampled, a useful range is, according to the gensim framework, (0, 1e-5).-window <window_size>
(default:5
)
The size of the window in the training process.-epochs <number_of_epochs>
(default:5
)
The number of epochs to use in training.-port <port_number>
(default:1808
)
The port that shall be used for the server.
Advanced Parameters
-continue <existing_walk_directory>
In some cases, old walks need to be re-used (e.g. if the program was interrupted after 48h). With the-continue
option, the walk generation can be continued; this means that old walks will be re-used and only missing walks are generated. This does not work for MID_WALKS (and flavors). If you do not need to generate additional walks use-onlyTraining
instead.
Besides generating walks and training embeddings, the CLI offers additional services which are described below.
(1) Full Vocabulary
jRDF2vec is compatible with the evaluation
framework for KG embeddings (GEval).
The latter framework requires the vectors to be present in a text file. If you have a gensim model or vector file,
you can use the following command to generate this file:
java -jar jrdf2vec-1.1-SNAPSHOT.jar -generateTextVectorFile ./path-to-your-model-or-vector-file
You can find the file (named vectors.txt
) in the directory where the model/vector file is located.
If you want to specify the file name/path yourself, you can use option -newFile <file_path>
.
(2) Subset of the Vocabulary
If you want to write a vectors.txt
file that contains only a subset of the vocabulary, you can additionally
specify the entities of interest using the -light <entity_file>
option (The <entity_file>
should contain one entity
(full URI) per line.):
java -jar jrdf2vec-1.1-SNAPSHOT.jar -generateTextVectorFile ./path-to-your-model-or-vector-file -light ./path-to-entity-file
You can find the file (named vectors.txt
) in the directory where the model/vector file is located.
If you want to specify the file name/path yourself, you can use option -newFile <file_path>
.
If the vector concepts contain surrounding tags that you want to remove in the process, use option -noTags
.
This command also works if ./path-to-your-model-or-vector-file
is an existing vector text file that shall be reduced.
jRDF2vec provides functionality to print all concepts for which a vector has been trained.
One word of the vocabulary will be printed per line to a file named vocabulary.txt
.
The model or vector file needs to be specified. If you have a gensim model or vector file, you can
use the following command to generate this file:
java -jar jrdf2vec-1.1-SNAPSHOT.jar -generateVocabularyFile ./path-to-your-model-or-vector-file
jRDF2vec generates a vectors.txt
file where one line represents a vector. This is the format also used by
GloVe, for instance.
In some cases, however, other file formats are required. You can use jRDF2vec to convert text vector files to other
common formats. The vector file does not have to be generated by jRDF2vec.
(1) Converting to w2v Format
To create a word2vec formatted file from the text file, you can use the following command:
java -jar jrdf2vec-1.1-SNAPSHOT.jar -convertToW2V <txt_file_path> <new_file.w2v>
(2) Converting to kv Format
The provided txt file (first parameter) can be either in txt
format or in w2v
format. Make sure you use the
correct file ending (.txt
/.w2v
).
You can run the command as follows:
java -jar jrdf2vec-1.1-SNAPSHOT.jar -convertToKv <txt_file_path> <new_file.kv>
(3) Converting to Tensorflow Projector Format
If you want to visualize your embedding space by using the Tensorflow Projector,
you can do so by converting your vectors.txt
file to the two files required by the tool. Use the following command:
java -jar jrdf2vec-1.1-SNAPSHOT.jar -convertToTfProjector <txt_file_path> [<vectors.tsv> <metadata.tsv>]
Two additional .tsv
files will be generated. You can find them in the same directory where <txt_file_path>
is
located.
Optionally, you can specify the paths of the files to be written as indicated in the command above.
For RDF2Vec, it is not always guaranteed that all concepts in the graph appear in the embedding space. For example,
some concepts may only appear in the object position of statements and may never be reached by random walks.
In addition, the word2vec configuration parameters may filter out infrequent words depending on the configuration (see
-minCount
above, for example). To analyze such rather seldom cases, you can use the -analyzeVocab
function specified
as follows:
java -jar jrdf2vec-1.1-SNAPSHOT.jar -analyzeVocab <model> <training_file|entity_file>
-
<model>
refers to any model representation such as gensim model file,.kv
file, or.txt
file. Just make sure you use the correct file endings. -
<training_file|entity_file>
refers either to the NT/TTL etc. file that has been used to train the model or to a text file containing the concepts you want to check (one concept per line in the text file, make sure the file ending is.txt
).
A report will be printed. For large models, you may want to redirect that into a file ([...] &> somefile.txt)
.
By default, jRDF2vec serializes walks in different gzipped files. If you require a single,
uncompressed file, you can use the -mergeWalks
keyword. You need to provide a
-walkDirectory <dir>
and you can optionally specify the output file using -o <file_path>
.
(Files not ending with .gz
in <dir>
will be skipped.)
java -jar jrdf2vec-1.1-SNAPSHOT.jar -mergeWalks -walkDirectory <dir> -o <file_to_write>
(1) Converting to PajekNet
To create a graph file in the PajekNet format (e.g. for graph analysis), you can use the following command:
java -jar jrdf2vec-1.1-SNAPSHOT.jar -convertToPajek <graph> <file_to_write>
Stable releases are available through the maven central repository:
<dependency>
<groupId>de.uni-mannheim.informatik.dws</groupId>
<artifactId>jrdf2vec</artifactId>
<version>1.0</version>
</dependency>
Optionally, Docker can be used to run jRDF2Vec. This functionality has been added by Vincent Emonet.
The Docker image can be used with the same arguments as the Jar file, refer to the documentation above for more details on the different jRDF2Vec arguments.
Test run to get the help message:
docker run -it --rm ghcr.io/dwslab/jrdf2vec -help
The best way to mount your local files in the docker container is to mount a folder on /data
in the container:
- On Linux and MacOS: use
$(pwd)
to mount the current working directory - On Windows: use
${PWD}
to mount the current working directory (and make the command in one line)
Here is an example generating embeddings using sample config files for DBpedia found in src/test/resources
in this repository. Use this command from the root folder of this repository on Linux or MacOS, change the $(pwd)
to ${PWD}
for Windows:
docker run -it --rm \
-v $(pwd):/data \
ghcr.io/dwslab/jrdf2vec \
-light /data/src/test/resources/sample_dbpedia_entity_file.txt \
-graph /data/src/test/resources/sample_dbpedia_nt_file.nt
Embeddings will be generated in the folders
walks
andpython_server
from where you ran the command.
A new docker image is automatically built and published to the GitHub Container Registry by a GitHub Actions workflow:
- The
latest
image tag is updated everytime a commit is pushed to themaster
branch - A new image tag is created for every new release published following the scheme
v0.0.0
Build from source code:
docker build -t ghcr.io/dwslab/jrdf2vec .
The most recent JavaDoc sites generated from the latest commit can be found here.
The following steps are necessary to obtain ordered RDF2vec embeddings (see publication Putting RDF2vec in Order for conceptional details).
Step 1: Generate Walks
Run jRDF2Vec to generate only walks (option -onlyWalks
) on your desired dataset.
Step 2: Merge the Walks in a single, uncompressed file
By default, jRDF2Vec serializes the walks in multiple gzipped files. For this application, however, we need a single,
uncompressed walk file.
You can use the corresponding jRDF2Vec command line service to do so.
Step 3: Compile wang2vec
Download the C implementation of wang2vec from GitHub.
Compile the files with make
.
Step 4: Run and have fun
Run the compiled wang2vec implementation on the merged walk file from step 2. In case you receive a segfault
error,
set the capping parameter to 1 (-cap 1
).
Call Syntax
./word2vec -train <your walk file> -output <desired file to be written> -type <2 (cwindow) or 3 (structured
skipgram>) -size <vector size> -threads <number of threads> -min-count 0 -cap 1
Exemplary Call
./word2vec -train walks.txt -output v100.txt -type 3 -size 100 -threads 4 -min-count 0 -cap 1
Not working? Contact us or open an issue.
Please do not forget to cite the corresponding papers:
(1) Portisch, Jan; Paulheim, Heiko. Putting RDF2vec in Order. In: Proceedings of the International Semantic Web
Conference - Posters and Demos, ISWC 2021. 2021.
(2) Ling, Wang; Dyer, Chris; Black, Alan; Trancoso, Isabel. Two/too simple adaptations of word2vec for syntax
problems. In: NAACL HLT 2015. pp. 1299–1304. ACL (2015)
I have Python installed, but it is not accessible via command python
. How to resolve this?
Create a file python_command.txt
in directory ./python_server
(created when first running the jar). Write the command
to call Python 3 in the first line of the file.
The program starts and immediately shuts down. Nothing seems to happen.
Make sure your system is set-up correctly, in particular whether you have installed Python 3 and the required
dependencies.
Can I run the command multiple times in parallel on the same machine?
Yes, you can. You need to make sure that for each command, you use (1) a different -port
and (2) a different
-walkDirectory
.