def get_ood_data(templates, train_dev_partition, test_partition, fo_test_ovit,
fo_test_ovot, output_header):
set_vocab_by_type('ood')
print('ood')
num_examples = 300
output_rows = []
guid = 0
for t in templates:
existing_templates = []
for i in range(num_examples):
generated_template = t.generate_one_example()
# check generated_template duplicate with existing
while generated_template in existing_templates:
# print('true')
generated_template = t.generate_one_example()
# add guid
output_rows.append([guid] + generated_template)
guid += 1
existing_templates.append(generated_template)
# filter by templates partitions
output_rows_test_ovit = filter_by_template_partition(
train_dev_partition, output_rows)
output_rows_test_ovot = filter_by_template_partition(
test_partition, output_rows)
write_csv(fo_test_ovit, output_rows_test_ovit, output_header)
write_csv(fo_test_ovot, output_rows_test_ovot, output_header)
def main():
hans_found_support_case = './esnli_train_subseq.csv'
output_file = './esnli_train_subseq_with_templates.csv'
output_header = None
output_rows = []
hans_templates = get_hans_templates()
with open(hans_found_support_case) as f:
reader = csv.reader(f)
for (i, line) in enumerate(reader):
if i == 0:
output_header = line
continue
label = line[1]
premise_pos = get_pos_tags(line[2])
hypothesis_pos = get_pos_tags(line[3])
template_name = match_templates(label, premise_pos, hypothesis_pos,
hans_templates)
if template_name != None:
line[-1] = template_name
print('template_name: ', template_name)
output_rows.append(line)
write_csv(output_file, output_rows, output_header)
def main():
fi = './templates.csv'
fo = './templates_new.csv'
rows = []
with open(fi) as f:
reader = csv.reader(f)
for (i, line) in enumerate(reader):
if i == 0:
header = line
if i > 0:
for j in range(len(line)):
item = line[j]
while item[0] == ' ':
item = item[1:]
while item[-1] == ' ':
item = item[:-1]
line[j] = item
rows.append(line)
write_csv(fo, rows, header)
def main():
data_dir_name = 'split_abundant_words_subcases'
num_seeds = 1
fi_name_list = ['dev_1', 'dev_2', 'dev_4', 'dev_7', 'dev_13', 'dev_32',
'train_1', 'train_2', 'train_4', 'train_8', 'train_16', 'train_32', 'train_64',
'test_ivit_300', 'test_ivot_300', 'test_ovit_300', 'test_ovot_300']
for seed in range(num_seeds):
for partition in range(5):
path = './%s/seed%d/partition%d/' % (data_dir_name, seed, partition)
for fi in fi_name_list:
fi_path = path + fi + '.csv'
fo_nl = path + fi + '_nl.csv' # natural language explanation
fo_pt = path + fi + '_pt.csv'# pointer-only explanations
fo_empty_expl = path + fi + '_empty_expl.csv' # empty_expl
nl_rows = []
pt_rows = []
empty_expl_rows = []
with open(fi_path) as f:
reader = csv.reader(f)
for (i, line) in enumerate(reader):
if i > 0:
guid = line[0]
label = line[5]
p = line[-4]
h = line[-3]
nl = line[-2]
pt = line[-1]
nl_row = [""]*19
nl_row[0] = guid
nl_row[1] = label
nl_row[2] = p
nl_row[3] = h
nl_row[4] = nl
pt_row = [""]*19
pt_row[0] = guid
pt_row[1] = label
pt_row[2] = p
pt_row[3] = h
pt_row[4] = pt
empty_expl_row = [""]*19
empty_expl_row[0] = guid
empty_expl_row[1] = label
empty_expl_row[2] = p
empty_expl_row[3] = h
nl_rows.append(nl_row)
pt_rows.append(pt_row)
empty_expl_rows.append(empty_expl_row)
write_csv(fo_nl, nl_rows, esnli_format_header)
write_csv(fo_pt, pt_rows, esnli_format_header)
write_csv(fo_empty_expl, empty_expl_rows, esnli_format_header)
def get_ind_data(templates, train_dev_partition, test_partition, fo_dir,
fo_train, fo_dev, fo_test_ivit, fo_test_ivot, output_header):
set_vocab_by_type('ind')
print('ind')
num_examples = 492
output_rows_train = []
output_rows_dev = []
output_rows_test = []
guid_train = 0
guid_dev = 0
guid_test = 0
output_rows_train_by_template = {}
output_rows_dev_by_template = {}
for t in templates:
existing_templates = []
for i in range(num_examples):
generated_template = t.generate_one_example()
# check generated_template duplicate with existing
while generated_template in existing_templates:
# print('true')
generated_template = t.generate_one_example()
# add guid
if i < 160:
line = [guid_train] + generated_template
template_id = generated_template[0]
if template_id in output_rows_train_by_template:
output_rows_train_by_template[template_id].append(line)
else:
output_rows_train_by_template[template_id] = [line]
output_rows_train.append(line)
guid_train += 1
elif i < 192:
line = [guid_dev] + generated_template
template_id = generated_template[0]
if template_id in output_rows_dev_by_template:
output_rows_dev_by_template[template_id].append(line)
else:
output_rows_dev_by_template[template_id] = [line]
output_rows_dev.append(line)
guid_dev += 1
else:
output_rows_test.append([guid_test] + generated_template)
guid_test += 1
existing_templates.append(generated_template)
# filter by templates partitions
output_rows_train_it = filter_by_template_partition(
train_dev_partition, output_rows_train)
output_rows_dev_it = filter_by_template_partition(train_dev_partition,
output_rows_dev)
output_rows_test_ivit = filter_by_template_partition(
train_dev_partition, output_rows_test)
output_rows_test_ivot = filter_by_template_partition(
test_partition, output_rows_test)
write_csv(fo_train, output_rows_train_it, output_header)
write_csv(fo_dev, output_rows_dev_it, output_header)
write_csv(fo_test_ivit, output_rows_test_ivit, output_header)
write_csv(fo_test_ivot, output_rows_test_ivot, output_header)
# few sample for train and dev sets
train_sample_sizes = [1, 2, 4, 8, 16, 32, 64]
dev_sample_sizes = list(set([int(0.2 * k) + 1
for k in train_sample_sizes]))
for train_size in train_sample_sizes:
fo_train = '%strain_%d.csv' % (fo_dir, train_size)
output_rows = []
for k, v in output_rows_train_by_template.items():
output_rows.extend(v[:train_size])
output_rows = filter_by_template_partition(train_dev_partition,
output_rows)
write_csv(fo_train, output_rows, output_header)
for dev_size in dev_sample_sizes:
fo_dev = '%sdev_%d.csv' % (fo_dir, dev_size)
output_rows = []
for k, v in output_rows_dev_by_template.items():
output_rows.extend(v[:dev_size])
output_rows = filter_by_template_partition(train_dev_partition,
output_rows)
write_csv(fo_dev, output_rows, output_header)
hyp_words = []
for word in premise.split():
if word not in [".", "?", "!"]:
prem_words.append(word.lower())
for word in hypothesis.split():
if word not in [".", "?", "!"]:
hyp_words.append(word.lower())
prem_filtered = " ".join(prem_words)
hyp_filtered = " ".join(hyp_words)
hypo_len = len(hypothesis.strip().split(" "))
expl1_len = len(expl1.strip().split(" "))
if hypo_len >= 3 and expl1_len >= 3:
if hyp_filtered in prem_filtered:
if label == "entailment":
count_entailment += 1
if label == "neutral":
count_neutral += 1
if label == "contradiction":
count_contradiction += 1
fo_row_data.append([guid, label, premise, hypothesis, expl1, ""])
print("Entailment:", count_entailment)
print("Contradiction:", count_contradiction)
print("Neutral:", count_neutral)
write_csv(fo_path, fo_row_data, fo_header)
