def _precision(self, ref, out, src, n):
"""
Calcualte GLEU-specific n-gram precision
Args:
ref: A reference sentence
out: An output sentence
src: A source sentence
Returns:
Numerator and denominator of the precision
"""
ref_ngram = ngram_utils.sent_ngrams_list(ref, n)
out_ngram = ngram_utils.sent_ngrams_list(out, n)
src_ngram = ngram_utils.sent_ngrams_list(src, n)
ref_cnt = Counter(ref_ngram)
out_cnt = Counter(out_ngram)
src_cnt = Counter(src_ngram)
out_join_ref = out_cnt & ref_cnt
out_join_src = out_cnt & src_cnt
num = sum(out_join_ref.values()) - \
sum((out_join_src - out_join_ref).values())
# According to https://github.com/cnap/gec-ranking/blob/master/scripts/gleu.py
num = max(num, 0)
denom = sum(out_cnt.values())
return num, denom
def num_repetitions_in_sentence(sentence: Tokens,
adjacent: bool = True,
ngram_order: int = 1) -> int:
"""
Counts repetitions in an input sentence.
:param sentence: A list of tokens or characters as strings.
:param adjacent: Whether repeated elements need to occur adjacent
to each other to count towards repetitions.
:param ngram_order: Order of ngrams considered, positive integer.
:return: Number of times an element was repeated.
"""
num_repetitions = 0
ngrams = ngram_utils.sent_ngrams_list(sentence, ngram_order)
if not adjacent:
counter = Counter(ngrams)
for k, v in counter.items():
num_repetitions += (v - 1)
else:
previous = []
for ngram in ngrams:
if ngram in previous:
num_repetitions += 1
previous.append(ngram)
if len(previous) > ngram_order:
previous.pop(0)
return num_repetitions
def _precision(self, ref, out, n):
"""
Caculate n-gram precision
Args:
ref: A reference sentence
out: An output sentence
Returns:
Numerator and denominator of the precision
"""
out_ngram = ngram_utils.sent_ngrams_list(out, n)
ref_ngram = ngram_utils.sent_ngrams_list(ref, n)
out_cnt = Counter(out_ngram)
ref_cnt = Counter(ref_ngram)
num = 0
denom = 0
for ngram, o_cnt in out_cnt.items():
num += min(o_cnt, ref_cnt[ngram])
denom += o_cnt
denom = max(1, denom)
return num, denom