gonenhila / gender_bias_lipstick Goto Github PK

In both remaining_bias_2016.ipynb and remaining_bias_2018.ipynb notebooks:

compute_corr(tuples_bef, 1, 3)
compute_corr(tuples_aft, 1, 3)

Should be changed to:

compute_corr(tuples_bef, 1, 3)
compute_corr(tuples_aft, 2, 3)

because tupple stores the following: (w, gender_bias_bef[w], gender_bias_aft[w], m, f)
while computing the correlation for the debiased embeddings, gender_bias_aft[w] which is at index 2 and not 1 should be used.
Also

def bias_by_neighbors(space, neighbours_num = 100):
    
    tuples = []
    for w in tqdm(vocab[space]):
        
        top = topK(w, space, k=neighbours_num+5)[:neighbours_num]

        m = 0
        f = 0    
        for t in top:
            if gender_bias_bef[t] > 0:
                m+=1
            else:
                f+=1
            
        tuples.append((w, gender_bias_bef[w], gender_bias_aft[w], m, f))

    return tuples
        

while checking for tuples_aft it should be changed to:

def bias_by_neighbors(space, neighbours_num = 100):
    
    tuples = []
    for w in tqdm(vocab[space]):
        
        top = topK(w, space, k=neighbours_num+5)[:neighbours_num]

        m = 0
        f = 0    
        for t in top:
            if gender_bias_aft[t] > 0: #This should be gender_bias_aft since we are calculating correlation in embedding after debiasing
                m+=1
            else:
                f+=1
            
        tuples.append((w, gender_bias_bef[w], gender_bias_aft[w], m, f))

    return tuples
   

Similar changes need to be made in the get_tuples_prof function under the title - Professions experiment.
Also,
For the Professions Expirements, I believe the following code:

tuples_bef_prof = get_tuples_prof('limit_bef', professions, gender_bias_bef)
tuples_aft_prof = get_tuples_prof('limit_aft', professions, gender_bias_bef)

should be changed to:

tuples_bef_prof = get_tuples_prof('limit_bef', professions, gender_bias_bef)
tuples_aft_prof = get_tuples_prof('limit_aft', professions, gender_bias_aft) 

I find that in most of the experiments, the gender bias by projection is mostly taken only for the original embeddings and almost never for the new ones.
Please let me know if I misunderstood.

EDIT - I misunderstood things. There is no problem with the code.

Dear authors,

I run your "remaining_bias_2016" notebook to check the clustering results. I notice that with your default random state setting, the clustering accuracy is 0.999 and 0.922 on my machine, which aligns with your paper report. But if I change to any other number, this result can significantly change.

For example, I can get original w2v with random state 0, resulting in clustering "precision" (accuracy) 0.001, which means it is not biased at all.

Or, I can get hard de-biased word embedding with clustering accuracy 0.075 with random state 2023, which means bias does not exist any more after hard debiasing.

Do you think this can significantly change some of your paper conclusions?