Debiasing-Multilingual-Word-Embeddings-A-Case-Study-of-Three-Indian-Languages

Here we have the code and data for the following paper: Debiasing Multilingual Word Embeddings: A Case Study of Three Indian Languages by Srijan Bansal, Vishal Garimella, Ayush Suhane, Animesh Mukherjee. Proceedings of ACM HyperText 2021

Setup Enviroment

• Create a conda environment with python 3.7.11
• Install requirments using: pip install -r requirements.txt inside the conda env

Download data

run bash download_embedding.sh The embeddings will be saved in data/embedding/ folder Download options:

• All: All the datasets, embeddings will be generated
• Essential: Fasttext embeddings, aligned text files, MKB datasets will be downloaded. Rest of the embeddings can be generated using scripts
• MKB: contains language pairs
• MKB_pickle: contains MKB dataset
• aligned_files: contains aligned files
• Bilingual: bilingual alinged embeddings (can be generated by generate_cross_lingual_embeddings.py)
• EQR: debiased embeddings (can be generated by running debias_run.py)
• LID: debiased embeddings (can be generated by running debias_run.py)

Scripts

• debias_run_runner.py: wrapper code for debias_run.py.
  • Run using python debias_run_runner.py
• debias_run.py: Generates debiased embeddings
  • Run using python debias_run.py lang1 lang2 pre num_lang_pairs
  • ex: python debias_run.py en hi LID 10
• generate_cross_lingual_embeddings_runner.py: wrapper code for generate_cross_lingual_embeddings.py
  • Run using python generate_cross_lingual_embeddings.py
• generate_cross_lingual_embeddings.py: Aligns embedding from language lang1 to lang2
  • Run using python generate_cross_lingual_embeddings.py lang1 lang2
  • ex: python generate_cross_lingual_embeddings.py bn hi
• MKB/alignment_runner.py: wrapper code for alignment.py
  • Run using python MKB/alignment.py
• MKB/alignment.py: Evaluates debiased embeddings
  • Run using python alignment.py pre1 pre2 lang1 lang2
  • ex: python alignment.py fin fin bn en
• MKB/bilingual_dict.py:
  • Generate pkl files. It might throw error due to problem with googletranslator api. Instead pkl files can be downloaded using MKB_pickle from download_embedding.sh

How to run

• Download datasets and embeddings using: download_embedding.sh with essential option
• Run generate_cross_lingual_embeddings_runner.py to generate all bilingual embeddings.
  • Input: MKD dataset, aligned text files from data/MKB and data/aligned_files
  • Output: Aligned embeddings will be saved to data/embedding/Bilingual/
• Run debias_run_runner.py to generate all debiased embeddings.
  • Input: Bilingual embeddings from data/embedding/Bilingual/
  • Output: Debiased embeddings will be saved to data/embedding/LID/, data/embedding/EQR/ etc
• Run MKB/alignment_runner.py to evaluate the debiased embeddings. Results will be saved to MKB/Alignment_runner.csv
  • Input: Debiased embeddings from data/embedding/LID/, data/embedding/EQR/ etc
  • Output: Debiased embeddings score will be saved to Alignment_Runner.csv