Repository files navigation

Implementation of translation-based relational learning models (TransE [Bordes et al. 2013], TransH) in Tensorflow.

In addition to original source code:More efficient triple scoring and mini-batch processing (Adagrad SGD).

Data preparation works directly with ontologies (RDF or OWL)For triple processing, rdflib is used.

For the use of sequential data (e.g. text corpora, events, etc.),there is the option to use a joint model (i.e. TransE + Skipgram (TEKE) either pre-trained or joint training)

More advanced sequential embedding models (event models) can be plugged-in (Convolutional-Autoencoder, Concatenation, ...)

Put thepath_to_kg and optionalpath_to_sequence inekl_experiment.py

Invoke: >python ekl_experiment.py

• Put an rdf/xml file into yourpath_to_kg
• Put asequence.txt file of comma-separated event IDs intopath_to_sequence
• Supply aunique_msgs.txt mapping of the form: Event-URI fragment identifier | ID (starting from 0)

• rdflib (4.1.2)
• pandas (0.19.2)
• numpy (1.13.0)
• TensorFlow (1.1.0)