def score_article(self,test,ref):
refs = [bleu.cook_refs([refSent],self.options['bleu_ngrams']) for refSent in ref]
testcook = []
for i,line in enumerate(test):
testcook.append(bleu.cook_test(line,refs[i],self.options['bleu_ngrams']))
score = bleu.score_cooked(testcook,self.options['bleu_ngrams'])
return score
def score_article(self, test, ref):
refs = [bleu.cook_refs([refSent], self.options["bleu_ngrams"]) for refSent in ref]
testcook = []
for i, line in enumerate(test):
testcook.append(bleu.cook_test(line, refs[i], self.options["bleu_ngrams"]))
score = bleu.score_cooked(testcook, self.options["bleu_ngrams"])
return score
def eval_sents(self,translist,targetlist):
scoredict = {}
cooked_test = {}
cooked_test2 = {}
cooktarget = [(items[0],bleu.cook_refs([items[1]],self.options['bleu_ngrams'])) for items in enumerate(targetlist)]
cooktarget = [(refID,(reflens, refmaxcounts, set(refmaxcounts))) for (refID,(reflens, refmaxcounts)) in cooktarget]
for testID,testSent in enumerate(translist):
scorelist = []
#copied over from bleu.py to minimize redundancy
test_normalized = bleu.normalize(testSent)
cooked_test["testlen"] = len(test_normalized)
cooked_test["guess"] = [max(len(test_normalized)-k+1,0) for k in range(1,self.options['bleu_ngrams']+1)]
counts = bleu.count_ngrams(test_normalized, self.options['bleu_ngrams'])
#separate by n-gram length. if we have no matching bigrams, we don't have to compare unigrams
ngrams_sorted = dict([(x,set()) for x in range(self.options['bleu_ngrams'])])
for ngram in counts:
ngrams_sorted[len(ngram)-1].add(ngram)
for (refID,(reflens, refmaxcounts, refset)) in cooktarget:
ngrams_filtered = ngrams_sorted[self.options['bleu_ngrams']-1].intersection(refset)
if ngrams_filtered:
cooked_test["reflen"] = reflens[0]
cooked_test['correct'] = [0]*self.options['bleu_ngrams']
for ngram in ngrams_filtered:
cooked_test["correct"][self.options['bleu_ngrams']-1] += min(refmaxcounts[ngram], counts[ngram])
for order in range(self.options['bleu_ngrams']-1):
for ngram in ngrams_sorted[order].intersection(refset):
cooked_test["correct"][order] += min(refmaxcounts[ngram], counts[ngram])
#copied over from bleu.py to minimize redundancy
logbleu = 0.0
for k in range(self.options['bleu_ngrams']):
logbleu += math.log(cooked_test['correct'][k])-math.log(cooked_test['guess'][k])
logbleu /= self.options['bleu_ngrams']
logbleu += min(0,1-float(cooked_test['reflen'])/cooked_test['testlen'])
score = math.exp(logbleu)
if score > 0:
#calculate bleu score in reverse direction
cooked_test2["guess"] = [max(cooked_test['reflen']-k+1,0) for k in range(1,self.options['bleu_ngrams']+1)]
logbleu = 0.0
for k in range(self.options['bleu_ngrams']):
logbleu += math.log(cooked_test['correct'][k])-math.log(cooked_test2['guess'][k])
logbleu /= self.options['bleu_ngrams']
logbleu += min(0,1-float(cooked_test['testlen'])/cooked_test['reflen'])
score2 = math.exp(logbleu)
meanscore = (2*score*score2)/(score+score2)
scorelist.append((meanscore,refID,cooked_test['correct']))
scoredict[testID] = sorted(scorelist,key=itemgetter(0),reverse=True)[:self.options['maxalternatives']]
return scoredict
def eval_sents(self, translist, targetlist):
scoredict = {}
cooked_test = {}
cooked_test2 = {}
cooktarget = [
(items[0], bleu.cook_refs([items[1]], self.options["bleu_ngrams"])) for items in enumerate(targetlist)
]
cooktarget = [
(refID, (reflens, refmaxcounts, set(refmaxcounts))) for (refID, (reflens, refmaxcounts)) in cooktarget
]
for testID, testSent in enumerate(translist):
scorelist = []
# copied over from bleu.py to minimize redundancy
test_normalized = bleu.normalize(testSent)
cooked_test["testlen"] = len(test_normalized)
cooked_test["guess"] = [
max(len(test_normalized) - k + 1, 0) for k in range(1, self.options["bleu_ngrams"] + 1)
]
counts = bleu.count_ngrams(test_normalized, self.options["bleu_ngrams"])
# separate by n-gram length. if we have no matching bigrams, we don't have to compare unigrams
ngrams_sorted = dict([(x, set()) for x in range(self.options["bleu_ngrams"])])
for ngram in counts:
ngrams_sorted[len(ngram) - 1].add(ngram)
for (refID, (reflens, refmaxcounts, refset)) in cooktarget:
ngrams_filtered = ngrams_sorted[self.options["bleu_ngrams"] - 1].intersection(refset)
if ngrams_filtered:
cooked_test["reflen"] = reflens[0]
cooked_test["correct"] = [0] * self.options["bleu_ngrams"]
for ngram in ngrams_filtered:
cooked_test["correct"][self.options["bleu_ngrams"] - 1] += min(
refmaxcounts[ngram], counts[ngram]
)
for order in range(self.options["bleu_ngrams"] - 1):
for ngram in ngrams_sorted[order].intersection(refset):
cooked_test["correct"][order] += min(refmaxcounts[ngram], counts[ngram])
# copied over from bleu.py to minimize redundancy
logbleu = 0.0
for k in range(self.options["bleu_ngrams"]):
logbleu += math.log(cooked_test["correct"][k]) - math.log(cooked_test["guess"][k])
logbleu /= self.options["bleu_ngrams"]
logbleu += min(0, 1 - float(cooked_test["reflen"]) / cooked_test["testlen"])
score = math.exp(logbleu)
if score > 0:
# calculate bleu score in reverse direction
cooked_test2["guess"] = [
max(cooked_test["reflen"] - k + 1, 0) for k in range(1, self.options["bleu_ngrams"] + 1)
]
logbleu = 0.0
for k in range(self.options["bleu_ngrams"]):
logbleu += math.log(cooked_test["correct"][k]) - math.log(cooked_test2["guess"][k])
logbleu /= self.options["bleu_ngrams"]
logbleu += min(0, 1 - float(cooked_test["testlen"]) / cooked_test["reflen"])
score2 = math.exp(logbleu)
meanscore = (2 * score * score2) / (score + score2)
scorelist.append((meanscore, refID, cooked_test["correct"]))
scoredict[testID] = sorted(scorelist, key=itemgetter(0), reverse=True)[: self.options["maxalternatives"]]
return scoredict
