djinn-anthrope / fuzzy_embeddings

This is the official repository of Bhat et. al. (2020): Word Representation as a Tuple of Feature Probabilities; RepL4NLP @ ACL 2020.

Code for "Word Embeddings as Tuples of Feature Probabilities"

Authors

Siddharth Bhat, Alok Debnath, Souvik Banerjee, and Manish Shrivastava

Kohli Center on Intelligent Systems, International Institute of Information Technology, Hyderabad

{siddharth.bhat, alok.debnath, souvik.banerjee}@research.iiit.ac.in

[e-mail]

Requirements

Please download the text8 corpus for creating word embeddings if you would like to use those made using the word2vec source code. You can also use other pretrained word embeddings for this task.

To replicate similarity measurements, please use the SimLex-999 corpus.

Instructions

cd word2vec
make all
./word2vec -train <input file name> -output <output file name> -size 200 -window 5 -sample 1e-4 -negative 5 -hs 0 -binary 0 -cbow 1 -iter 3

To perform analogy, simply change the model path name in 1-eval.sh and run it as ./1-eval.sh

To perform similarity, simply change the model path name and simlex corpus path name in 1-paper-numbers.sh and run it as ./1-eval.sh

To use the command line interface for fuzzy embeddings, run:

make cli
./cli

and follow the instructions.