multilingual-knowledge-graph-bias: A Python repository from PootieT

Knowledge Graph Embedding Bias Analysis with Poincare and Euclidean embedding methods

Language Selection

For the DBpedia knowledge graphs we are looking into, we decided to pick English (en), Swedish (sv), and Indonesian (id).

We picked English because it is the most relevant language of choice. We picked Swedish because Scadanavian countries are known for exhibiting less gender inequality, and Swedish has the largest DBpedia collection out of all other Scandanavian countries (Swedish (221.1 Mb), Danish (20.1 Mb), Norwegian (32.6 Mb), Icelandic (2.6 Mb), and Faroese (496 Kb)). Similarly, we picked Indonesian because it is one of countries with relatively more orthodox traditions and has the most data (Indonesian (60.3 Mb), Urdu (21 Mb), Hindi (11.3 Mb), and Arabic (5.5 Mb))

Data Preprocessing

To run data preprocessing code, see in main functions in data_explorations.py, specifically, import_dbpedia_data for dbpedia datasets.

To train a model

To train a model, see code in run_train.py

Available datasets are:

FB15k-237
WN18RR
DBpedia-english
DBpedia-Swedish
DBpedia-Indonesian
(and other languages using `/download_datasets.sh` file

Our training statistics is automatically logged in WanDB. To install:

pip install wandb
wandb login

To Compute Bias

First you need three files for the dataset:

humans.ent
gender.ent
professions.ent

To generate those, run data_exploration.extract_dbpedia_entities

Then, see code in run_predict.py

Disclaimer