This repository contains code and data for the experiments in the paper:
Unsupervised Embedding Enhancements of Knowledge Graphs using Textual Associations
by Neil Veira, Brian Keng, Kanchana Padmanabhan, Andreas Veneris
Published at IJCAI 2019.
The code requires the following packages:
- Python 3.6
- numpy 1.15.4
- scipy 1.0.0
- scikit-learn 0.19.1
- gensim 3.5
- nltk 3.3
- tensorflow 1.5
All experimental configurations can be run through the script run.py (see run.py --help for details), which performs the data processing, training, and evaluation steps. Seven bash scripts are provided to reproduce the seven configurations from the paper. The command format is run_<model>.sh <dataset> <scoring function> .
Supported datasets include Wordnet (WN) and Freebase (FB)
Supported scoring functions include Structured Embeddings (SE), Translational Embeddings (TransE), Translational Relations (TransR), RESCAL (RESCAL), DistMult (DistMult), and Holographic Embeddings (HolE).
For example, to run the FeatureSum model on Wordnet using the DistMult scoring function, run
./run_FeatureSum.sh WN DistMult
Word embedding vectors learned by the WeightedWordVectors model for Wordnet and Freebase can be found in the trained_embeddings directory. Each file is a pickled dict mapping words (strings) to 100-dimensional numpy arrays.