RelSifter: Scoring Triples for Type-Like Relations

Team: samphire at WSDM Cup 2017 held at Cambridge, UK.

Pre-requisites

Download a .zip file from wsdm-cup-2017-models containing all the following:

• knowledge graph
• machine learned models for profession and nationality

Getting started

1. Clone this repository and cd into it.
2. Under the directory relsifter, place the uncompressed directory wsdm-cup-2017-models and rename it to model.

Installing RelSifter

Navigate to the root directory and run the following command. This may take a while.

    python setup.py install

Using RelSifter

Once installed, the following command can be used to run RelSifter. This will create an output file with the same name in the directory specified by the output flag.

    relsifter -i input.txt -o ./

Development mode

Start with installing RelSifter in development mode to experiment with extracting features and building models for predicting relevance scores for type-like relations.

    python setup.py develop

Generating features for learning

1. TF-IDF features: Navigate to relsifter/characterization and use the compute_pertinence.py module to compute combined pertinence.
2. Text based features: Navigate to relsifter/textprofile and use the feature_extraction.py module to compute Wikipedia abstracts-based features.

Building machine learning models

1. TF-IDF based model: Navigate to relsifter/characterization and use the model_building.py module to train RandomForest, Adaboost and/or Ordinal Logistic Regression.
2. Text based model: Navigate to relsifter/textprofile and use the model_building module to build RandomForest, Adaboost and/or Ordinal Logistic Regression.

Acknowledgments

Fabian Pedregosa-Izquierdo. Feature extraction and supervised learning on fMRI : from practice to theory. Medical Imaging. Université Pierre et Marie Curie - Paris VI, 2015. English. Github repository: mord