data_dir = os.path.join(MOVERSCORE_DIR, 'MT')
reference_list = dict({
"newstest2017-csen-ref.en": "cs-en",
# "newstest2017-deen-ref.en": "de-en",
# "newstest2017-ruen-ref.en": "ru-en",
# "newstest2017-tren-ref.en": "tr-en",
# "newstest2017-zhen-ref.en": "zh-en"
})
metric = 'MoverScore'
data = []
for _ in reference_list.items():
reference_path, lp = _
references = load_data(os.path.join(data_dir, reference_path))
with MosesDetokenizer('en') as detokenize:
references = [detokenize(ref.split(' ')) for ref in references]
idf_dict_ref = get_idf_dict(references)
all_meta_data = load_metadata(os.path.join(data_dir, lp))
for i in tqdm.tqdm(range(len(all_meta_data))):
path, testset, lp, system = all_meta_data[i]
translations = load_data(path)
with MosesDetokenizer('en') as detokenize:
translations = [detokenize(hyp.split(' ')) for hyp in translations]
idf_dict_hyp = get_idf_dict(translations)
df_system = pd.DataFrame(columns=('metric', 'lp', 'testset', 'system',
'sid', 'score'))
scores = word_mover_score(references,
import torch
import truecase
from mosestokenizer import MosesDetokenizer
from mt_utils import (find_corpus, load_data, load_metadata,
print_sys_level_correlation, print_seg_level_correlation,
df_append)
dataset = find_corpus(args.dataset)
wmt_xmoverscores = []
for pair in dataset.items():
reference_path, lp = pair
src, tgt = lp.split('-')
references = load_data(
os.path.join(args.dataset, 'references/', reference_path))
source_path = reference_path.replace('ref', 'src')
source_path = source_path.split('.')[0] + '.' + src
source = load_data(os.path.join(args.dataset, 'source', source_path))
all_meta_data = load_metadata(
os.path.join(args.dataset, 'system-outputs', lp))
with MosesDetokenizer(src) as detokenize:
source = [detokenize(s.split(' ')) for s in source]
with MosesDetokenizer(tgt) as detokenize:
references = [detokenize(s.split(' ')) for s in references]
device = 'cuda:0'
temp = np.loadtxt('mapping/europarl-v7.' + src + '-' + tgt +