DS-20202's picture
The app
a1a1ad8
import urllib.request
import gradio as gr
import numpy as np
from utils import similarity
import scipy
def load_glove(path):
with open(path) as f:
lines = f.readlines()
wv = []
vocab = []
for line in lines:
tokens = line.strip().split(" ")
assert len(tokens) == 301
vocab.append(tokens[0])
wv.append([float(elem) for elem in tokens[1:]])
w2i = {w: i for i, w in enumerate(vocab)}
wv = np.array(wv).astype(float)
print(len(vocab), wv.shape, len(w2i))
return wv, w2i, vocab
urllib.request.urlretrieve("https://cdn-lfs.huggingface.co/repos/51/d0/51d02f0735de2187e78af7db593c8d71efbcddfe9f5e45cfa1c77904daf0dfdf/20dac974c73413d4b4f463f9b28eabfb326793d2732c108a9cd1262f2a055dc3?response-content-disposition=attachment%3B%20filename%3D%22glove_debiased_300d.txt%22", "glove_debiased_300d.txt")
urllib.request.urlretrieve("https://cdn-lfs.huggingface.co/repos/51/d0/51d02f0735de2187e78af7db593c8d71efbcddfe9f5e45cfa1c77904daf0dfdf/91125602f730fea7ca768736c6f442e668b49db095682bf2aad375db061c21ed?response-content-disposition=attachment%3B%20filename%3D%22glove.6B.300d.txt%222", "glove_300d.txt")
original_wv, original_w2i, original_vocab = load_glove('glove_300d.txt')
wv_debiased, w2i_debiased, vocab_debiased = load_glove('glove_debiased_300d.txt')
def debias(gendered_word1,gendered_word2,occupation,model):
# if model == 'Glove-300d' :
# original_wv, original_w2i, original_vocab = load_glove('glove_300d.txt')
# wv_debiased, w2i_debiased, vocab_debiased = load_glove('glove_debiased_300d.txt')
print(similarity('man', 'nurse', original_wv, original_w2i))
return abs(similarity(gendered_word1, occupation, wv_debiased, w2i_debiased)-similarity(gendered_word2, occupation, wv_debiased, w2i_debiased)), \
abs(similarity(gendered_word1, occupation, original_wv, original_w2i)-similarity(gendered_word2, occupation, original_wv, original_w2i)), "<h3>How the difference is measured?</h3> <p>abs(similarity(gendered_word1,occupation)-similarity(gendered_word2,occupation))</p>"
demo = gr.Interface(
debias,
inputs = [
gr.Textbox(placeholder="Gendered Word 1(Exp. man)"),
gr.Textbox(placeholder="Gendered Word 2(Exp. woman)"),
gr.Textbox(placeholder="Occupation(Exp. nurse)"),
gr.Radio(choices=['Glove-300d'],value='Glove-300d')
],
outputs = [gr.Textbox(label="Absolute similarity differece in the debiased version"),gr.Textbox(label="Absolute similarity differece in the original version"), gr.Markdown(
value="<h3>How the difference is measured?</h3> <p>abs(similarity(gendered_word1,occupation)-similarity(gendered_word2,occupation))</p>")],
description = '<a href="https://aclanthology.org/2020.acl-main.484/">Double-Hard Debias: Tailoring Word Embeddings for Gender Bias Mitigation</a>'
)
if __name__ == '__main__':
demo.launch()