def compare_with_random(load_emb):
pkl_path = './data-toxic-kaggle/toxic_comments_100.pkl'
perturbed_path = './data-toxic-kaggle/toxic_comments_100_perturbed.pkl' # perturbed by Edwin's script
original, perturbed = data_helper.load_samples(perturbed_path,pkl_path)
logging.info("data loaded")
perturbed = original[:]
np.random.shuffle(perturbed)
logging.info("sentences are shuffled")
output = []
for index in range(len(original)):
logging.info("index:%d"%index)
original_sample = original[index]
perturbed_sample = perturbed[index]
#print([original_sample, perturbed_sample])
(vec_orig, vec_pert) = embeddings.get_embeddings([original_sample, perturbed_sample],load_emb)
output.append("%s,%s,%.2f"%(original_sample,perturbed_sample, cos_sim(vec_orig, vec_pert)))
return output
def compare_with_probaility(load_emb, perturbation_script, p=1.0):
pkl_path = './data-toxic-kaggle/toxic_comments_100.pkl'
if perturbation_script == 'steffen_even':
perturbed_path = './data-toxic-kaggle/toxic_comments_100_mm_even_p%.1f.pkl'%p # perturbed by Steffen_even script
original, perturbed = data_helper.load_samples(perturbed_path,pkl_path)
logging.info("perturbation_script: steffen_even")
logging.info("data loaded")
output = []
for index in range(len(original)):
logging.info("index:%d"%index)
original_sample = original[index]
perturbed_sample = perturbed[index]
(vec_orig, vec_pert) = embeddings.get_embeddings([original_sample, perturbed_sample],load_emb)
output.append("%s,%s,%.2f"%(original_sample,perturbed_sample, cos_sim(vec_orig, vec_pert)))
elif perturbation_script == 'edwin':
perturbed_path = './data-toxic-kaggle/toxic_comments_100_perturbed.pkl' # perturbed by Edwin's script
original, perturbed = data_helper.load_samples(perturbed_path,pkl_path)
logging.info("perturbation_script: edwin")
logging.info("data loaded")
output = []
for index in range(len(original)):
logging.info("index:%d"%index)
original_sample = original[index]
perturbed_sample = perturbed[index]
pert = ''
for i,ch in enumerate(original_sample):
prob = np.random.uniform()
if ( prob<= p) and (ch not in ignore_list):
# disturb
pert += perturbed_sample[i]
else:
pert += ch
#print([original_sample,pert])
(vec_orig, vec_pert) = embeddings.get_embeddings([original_sample, pert],load_emb)
output.append("%s,%s,%.2f"%(original_sample,pert, cos_sim(vec_orig, vec_pert)))
else:
raise ValueError("%s is not implemented!"%perturbation_script)
return output
def cosine_similarities(loaded_emb, clean_path, perturbed_path):
clean_sentences, perturbed_sentences = data_helper.load_samples_txt(
clean_path, perturbed_path)
n = len(clean_sentences)
cosine_similarities = []
for index in range(n):
clean_sentence = clean_sentences[index]
perturbed_sentence = perturbed_sentences[index]
(vec_clean, vec_pert) = embeddings.get_embeddings(
[clean_sentence, perturbed_sentence], loaded_emb)
cosine_similarities.append(cos_sim(vec_clean, vec_pert))
return (clean_sentences, perturbed_sentences, cosine_similarities)
def online_compare(load_emb, original, perturbed):
print("data loaded")
output = []
(vec_orig, vec_p_1) = embeddings.get_embeddings([original, perturbed],load_emb)
return cos_sim(vec_orig, vec_p_1)