swd543 / FaVLib

Fact Validation Library

FaVLib

In order to use an integrated solution to the fact validation problem, we developed a Fact Validation Library, called FaVLib, with a workflow implementation that uses Docker and Common Workflow Language. This library trains a classifier on embedding features to predict truthfulness of a given fact based on other facts already present in the knowledge graph.

The FaVLib uses Docker and Common Workflow Language (CWL) representation to properly configure software components and easily execute the workflow. The workflow depicted in Figure has three main steps:

1. data generation
2. embedding learning
3. triple classification

In the first step, the data generation is carried out on a Knowledge Graph (KG) and the parameters (e.g., negative sampling strategy and test fraction ratio) to generate the necessary data sets for embedding learning and machine learning steps. Next, in the embedding learning step, training positives are used to learn entity embeddings based on the configured embedding method and embedding parameters. In the final step, basic classifiers are trained on the training data and evaluated on test data using embedding features, and feature vectors are exported to be re-usable for other machine learning algorithms / platforms.

The library integrates various tools including AYNEC (https://github.com/tdg-seville/AYNEC), which generates negative/positive samples and splits data into train and test set, and PyKEEN (https://github.com/SmartDataAnalytics/PyKEEN), which learns multiple embedding methods and machine learning methods for triple fact classification.

How to run it

Using Docker

• First make sure Docker is installed! (see below)
• Clone this repository

git clone https://github.com/MaastrichtU-IDS/FaVLib.git

• Move to the cloned repository

cd FavLib

• Pull the container image (or build it if you prefer)

docker pull umids/favlib

• Run the container on port 8888

docker run -d  --rm --name favlib -p 8888:8888 -v $(pwd):/jupyter -v /tmp:/tmp umids/favlib

Access the Jupyter notebook on http://localhost:8888

• Run a workflow:

docker exec -it favlib cwltool --outdir=/jupyter/output/ workflow/main-workflow-pykeen.cwl workflow/workflow-pykeen.yml

• workflow/main-workflow-pykeen.cwl : workflow description defined in CWL
• workflow/workflow-pykeen.yml : configuration file, need to define your input and parameter
• --outdir=/jupyter/output/ : the output of the workflow will be stored in '/jupyter/output/'

Look at the configuration file, workflow/workflow-pykeen.yml :

Data Generation

• inputFile: /jupyter/data/input/sample_kg.tsv
• minNumRel: 1
• negStrategy: 'change_source'
• fractionTest: 0.1
• predict: 1 # to make prediction
• predicate: '<hasCapital>' # make prediction for only given predicate
• numTrainNegatives: 1 # number of negative for each triple in train
• numTestNegatives: 10 # number of negative for each triple in train

Embedding Learning

• embedding_model_name: 'TransE'
• embedding_dim: 50
• normalization_of_entities: 2
• seed: 0
• scoring_function: 1
• margin_loss: 1
• learning_rate: 0.01
• batch_size: 64
• num_epochs: 50
• filter_negative_triples: True
• preferred_device: 'cpu'

Triple Classification

• output_train: 0 # to output the training feature matrix (to test your with different model)

• output_test: 0 # to output the test feature matrix

• That's it!

Install Docker

https://docs.docker.com/install/

Windows requires to have:

• Windows 10 64-bit: Pro, Enterprise, or Education (Build 15063 or later).

• Hyper-V and Containers Windows features must be enabled.

See this page for some basic informations about how to run Docker, connect it to your laptop storage and deploy applications locally.