def test_bleu(self):
bleu = BLEUCalculator()
score = bleu.bleu(
"I am waiting on the beach",
"He is walking on the beach",
)
score_from_list = bleu.bleu("I am waiting on the beach".split(),
["He is walking on the beach".split()])
self.assertLess(abs(score - score_from_list), 1e-8)
bleu = BLEUCalculator(lang="ja")
score_ja = bleu.bleu("私はビーチで待ってる", "彼がベンチで待ってる")
self.assertLess(abs(score - score_ja), 1e-8)
def eval_rouges(refrence_summary, model_summary):
# refrence_summary = "tokyo shares close up #.## percent"
# model_summary = "tokyo stocks close up # percent to fresh record high"
rouge = RougeCalculator(stopwords=True, lang="en")
rouge_1 = rouge.rouge_n(
summary=model_summary,
references=refrence_summary,
n=1)
rouge_2 = rouge.rouge_n(
summary=model_summary,
references=[refrence_summary],
n=2)
rouge_l = rouge.rouge_l(
summary=model_summary,
references=[refrence_summary])
# You need spaCy to calculate ROUGE-BE
rouge_be = rouge.rouge_be(
summary=model_summary,
references=[refrence_summary])
bleu = BLEUCalculator()
bleu_score = bleu.bleu(summary=model_summary,
references=[refrence_summary])
# print("ROUGE-1: {}, ROUGE-2: {}, ROUGE-L: {}, ROUGE-BE: {}".format(
# rouge_1, rouge_2, rouge_l, rouge_be
# ).replace(", ", "\n"))
return rouge_1, rouge_2, rouge_l, rouge_be, bleu_score
def convert_excel_to_df_and_evaluate(self):
"""エクセルをDataFrameへ"""
input_df = pd.ExcelFile(self.input_file).parse()
bleu_ja = BLEUCalculator(lang="ja")
rouge = RougeCalculator(lang="ja")
output_dict = {
'bleu': [],
'rouge_1': [],
'rouge_2': [],
'rouge_long': []
}
for index, row in input_df.iterrows():
output_dict['bleu'].append(
bleu_ja.bleu(row['ref_text'], row['input_text'])) # BLEUでの評価
output_dict['rouge_1'].append(
rouge.rouge_n(summary=row['ref_text'],
references=row['input_text'],
n=1)) # ROUGE(n1)での評価
output_dict['rouge_2'].append(
rouge.rouge_n(summary=row['ref_text'],
references=row['input_text'],
n=2)) # ROUGE(n2)での評価
output_dict['rouge_long'].append(
rouge.rouge_l(
summary=row['ref_text'],
references=row['input_text'])) # ROUGE(rouge_l)での評価
# input_microphone_df['rouge_be'] = rouge.rouge_be(summary=row['ref_text'], references=row['input_text']) # ROUGE(rouge_be)での評価
input_df['bleu'] = output_dict['bleu']
input_df['rouge_1'] = output_dict['rouge_1']
input_df['rouge_2'] = output_dict['rouge_2']
input_df['rouge_long'] = output_dict['rouge_long']
# print(input_df)
return input_df
def main(args):
system_out = read_file(args.system_output)
reference_list = read_file(args.reference)
bleu = BLEUCalculator()
bleu_list = []
for index, snt in enumerate(system_out):
bleu_list.append(
bleu.bleu(summary=snt, references=reference_list[index]))
print('SACRE_BLEU\t%.6f' % (np.average(bleu_list)))
def evaluate_bleu(summary, references, lang="zh"):
bleu_calc = BLEUCalculator(lang=lang)
assert len(summary) == len(references), "number of summary and references should be equal"
scores = []
for s, rs in zip(summary, references):
score = bleu_calc.bleu(s, rs)
scores.append(score)
score_avg = sum(scores) / len(scores)
return score_avg, scores
def cal_bleu(prediction_str, target_str):
bleu = BLEUCalculator()
total_bleu = []
for index in range(len(prediction_str)):
prediction_rel = ' '.join(prediction_str[index])
eos_index = prediction_rel.find('<eos>')
if (eos_index > 0):
prediction_rel = prediction_rel[:eos_index - 1]
target_rel = ' '.join(target_str[index])
target_rel = target_rel[:target_rel.find('<eos>') - 1]
total_bleu.append(bleu.bleu(prediction_rel, target_rel))
return np.mean(total_bleu)
class LanguageMetrics(object):
bleu = BLEUCalculator(tokenizer=SimpleTokenizer())
rouge = RougeCalculator(stopwords=True,
lang="en",
tokenizer=SimpleTokenizer())
@staticmethod
def _computeScore(summary, refs, criteria):
if isinstance(refs, str):
refs = [refs]
score = criteria(summary=summary, references=refs)
return score
@staticmethod
def blue_score(summary, refs):
score = LanguageMetrics._computeScore(summary, refs,
LanguageMetrics.bleu.bleu)
return score
@staticmethod
def rouge_1_score(summary, refs):
score = LanguageMetrics._computeScore(summary, refs,
LanguageMetrics.rouge.rouge_1)
return score
@staticmethod
def rouge_2_score(summary, refs):
score = LanguageMetrics._computeScore(summary, refs,
LanguageMetrics.rouge.rouge_2)
return score
@staticmethod
def rouge_l_score(summary, refs):
score = LanguageMetrics._computeScore(summary, refs,
LanguageMetrics.rouge.rouge_l)
return score
@staticmethod
def rouge_be_score(summary, refs):
score = LanguageMetrics._computeScore(summary, refs,
LanguageMetrics.rouge.rouge_be)
return score
@staticmethod
def rouge_n_score(summary, refs, n):
rouge_n = partial(func=LanguageMetrics.rouge.rouge_n, n=n)
score = LanguageMetrics._computeScore(summary, refs, rouge_n)
return score
def myeval(valid_x, valid_y, vocab, model):
rouge = RougeCalculator(stopwords=True, lang="zh")
bleu_ch = BLEUCalculator(lang="zh")
model.eval()
eval_batch_num = 0
sum_rouge_1 = 0
sum_rouge_2 = 0
sum_rouge_L = 0
score_ch = 0
sum_loss = 0
limit = 63
logging.info('Evaluating on %d minibatches...' % limit)
i2w = {key: value for value, key in vocab.items()}
ckpt_file = args.ckpt_file[9:]
fout_pred = open(os.path.join('tmp/systems', '%s.txt' % ckpt_file), "w")
fout_y = open(os.path.join('tmp/models', 'ref_%s.txt' % ckpt_file), "w")
while eval_batch_num < limit:
with torch.no_grad():
loss = run_batch(valid_x, valid_y, model)
sum_loss += loss
_, x = valid_x.next_batch()
pred = greedy(model, x, vocab)
_, y = valid_y.next_batch()
y = y[:,1:].tolist()
for idx in range(len(pred)):
line_pred = [i2w[tok] for tok in pred[idx] if tok != vocab[config.end_tok] and tok != vocab[config.pad_tok]]
line_y = [i2w[tok] for tok in y[idx] if tok != vocab[config.end_tok] and tok != vocab[config.pad_tok]]
fout_pred.write(" ".join(line_pred) + "\n")
fout_y.write(" ".join(line_y) + "\n")
sum_rouge_1 += rouge.rouge_n(references=" ".join(line_y),summary=" ".join(line_pred),n=1)
sum_rouge_2 += rouge.rouge_n(references=" ".join(line_y),summary=" ".join(line_pred),n=2)
sum_rouge_L += rouge.rouge_l(references=" ".join(line_y),summary=" ".join(line_pred))
score_ch += bleu_ch.bleu(" ".join(line_y), " ".join(line_pred))
eval_batch_num += 1
fout_pred.close()
fout_y.close()
avg_rouge_1 = sum_rouge_1/(len(pred) * limit)
avg_rouge_2 = sum_rouge_2/(len(pred) * limit)
avg_rouge_L = sum_rouge_L/(len(pred) * limit)
avg_bleu_ch = score_ch/(len(pred) * limit)
avg_loss = sum_loss/limit
print("ROUGE_1 = ",avg_rouge_1)
print("ROUGE_2 = ",avg_rouge_2)
print("ROUGE_L = ",avg_rouge_L)
print("BLEU = ", avg_bleu_ch)
print("Perplexity = ", math.pow(2, avg_loss))
model.train()
def __init__(self,
metrics: List[str] = [
"rouge_1", "rouge_2", "rouge_l", "rouge_be", "bleu"
],
lang: str = "en",
stopwords: bool = True,
stemming: bool = True,
use_porter=True):
if use_porter:
self.rouge = RougeCalculator(stopwords=stopwords,
stemming=stemming,
lang="en-porter")
else:
self.rouge = RougeCalculator(stopwords=stopwords,
stemming=stemming,
lang="en")
self.bleu = BLEUCalculator(lang=lang)
self.metrics = sorted(metrics)
def computeSacreBleu(translation_path,
reference_path,
lang,
detokenize_trans=True,
detokenize_ref=False):
bleu = BLEUCalculator(lang=lang)
trans_raw = trans = readSentences(translation_path)
reference_raw = reference = readSentences(reference_path)
if detokenize_trans or detokenize_ref:
detok = MosesDetokenizer(lang)
if detokenize_trans:
trans = [detok([d]) for d in trans_raw]
if detokenize_ref:
reference = [detok([d]) for d in reference_raw]
bleu_score = bleu.bleu(summary=trans,
references=[reference],
score_only=True)
print(bleu_score)
return bleu_score
def test_custom_lang(self):
class Custom(BaseLang):
def __init__(self):
super(Custom, self).__init__("cs")
def tokenize(self, text):
return text.split("/")
lang = Custom()
rouge = RougeCalculator(lang=lang)
rouge_score = rouge.rouge_n(
summary="I/went/to/the/Mars/from/my/living/town.",
references="I/went/to/Mars",
n=1)
bleu = BLEUCalculator(lang=lang)
bleu_score = bleu.bleu("I/am/waiting/on/the/beach",
"He/is/walking/on/the/beach")
self.assertGreater(rouge_score, 0)
self.assertGreater(bleu_score, 0)
def __init__(self):
self.bleu_calc = BLEUCalculator()
def main(score_desc: ("ex: To calculate ROUGE-N, L, BE => 'r-nlb'"),
use_file: ("read data from file", "flag", "f"),
include_stopwords: ("don't ignore stop words", "flag", "in"),
stemming: ("use stemming", "flag", "st"),
word_limit: ("word limit count", "option", "wl") = -1,
length_limit: ("sentence limit length", "option", "ll") = -1,
alpha: ("alpha for f1-score", "option") = 0.5,
language: ("word limit count", "option", "la") = "en",
*params):
if "-" in score_desc:
score_type, score_kinds = score_desc.lower().split("-")
else:
score_type = score_desc.lower()
score_kinds = ""
if len(params) < 2:
print("You have to specify at least one summary and reference.")
return
summary = params[0]
references = params[1:]
if isinstance(references, tuple):
references = list(references)
stopwords = not include_stopwords
generator = None
if use_file:
generator = file_generator(summary, references)
else:
generator = sentence_to_generator(summary, references)
scores = []
keys = []
if score_type == "r":
scorer = RougeCalculator(stopwords=stopwords,
stemming=stemming,
word_limit=word_limit,
length_limit=length_limit,
lang=language)
for s, rs in generator:
score = {}
for k in score_kinds:
if k == "n":
score["ROUGE-1"] = scorer.rouge_1(s, rs, alpha)
score["ROUGE-2"] = scorer.rouge_2(s, rs, alpha)
elif k == "l":
score["ROUGE-L"] = scorer.rouge_l(s, rs, alpha)
elif k == "b":
score["ROUGE-BE"] = scorer.rouge_be(s, rs, "HMR", alpha)
if len(keys) == 0:
keys = list(score.keys())
scores.append(score)
elif score_type == "b":
scorer = BLEUCalculator(lang=language)
for s, rs in generator:
score = {}
print(s, rs)
score["BLEU"] = scorer.bleu(s, rs)
if len(keys) == 0:
keys = list(score.keys())
scores.append(score)
avgs = {}
for k in keys:
avg = mean([s[k] for s in scores])
avgs[k] = avg
result = {
"options": {
"stopwords": stopwords,
"stemming": stemming,
"word_limit": word_limit,
"length_limit": length_limit,
"alpha": alpha,
"input-summary": summary,
"input-references": references
},
"averages": avgs,
"scores": scores
}
output = json.dumps(result, indent=2, ensure_ascii=False)
print(output)
document.ref_keywords)
tokens: Set[str] = retrieve_lemmatized_tokens(document.lang,
document.keywords)
tp = len(ref_tokens.intersection(tokens))
fp = len(tokens) - tp
fn = len(ref_tokens) - tp
eps = 1e-6
precision = tp / (tp + fp + eps)
recall = tp / (tp + fn + eps)
f1 = 2 * (precision * recall) / (precision + recall + eps)
res = {'precision': precision, 'recall': recall, 'f1': f1}
return res
rouge_en = RougeCalculator(stopwords=True, lang="en")
bleu = BLEUCalculator()
def evaluate_summary(document: DocumentForEval) -> Dict[str, float]:
summary = '\n'.join(document.summary)
ref_summary = '\n'.join(document.ref_summary)
if document.lang.lower() != 'en':
raise ValueError("Only English language is supported at the moment.")
rouge_1 = rouge_en.rouge_n(summary=summary, references=ref_summary, n=1)
rouge_2 = rouge_en.rouge_n(summary=summary, references=ref_summary, n=2)
rouge_3 = rouge_en.rouge_n(summary=summary, references=ref_summary, n=3)
rouge_4 = rouge_en.rouge_n(summary=summary, references=ref_summary, n=4)
#Evaluation task ##################################
from sumeval.metrics.rouge import RougeCalculator
rouge = RougeCalculator(stopwords=True, lang="en")
rouge_1 = rouge.rouge_n(summary="I went to the Mars from my living town.",
references="I went to Mars",
n=1)
rouge_2 = rouge.rouge_n(summary="I went to the Mars from my living town.",
references=["I went to Mars", "It's my living town"],
n=2)
rouge_l = rouge.rouge_l(summary="I went to the Mars from my living town.",
references=["I went to Mars", "It's my living town"])
# You need spaCy to calculate ROUGE-BE
rouge_be = rouge.rouge_be(summary="I went to the Mars from my living town.",
references=["I went to Mars", "It's my living town"])
print("ROUGE-1: {}, ROUGE-2: {}, ROUGE-L: {}, ROUGE-BE: {}".format(
rouge_1, rouge_2, rouge_l, rouge_be).replace(", ", "\n"))
from sumeval.metrics.bleu import BLEUCalculator
bleu = BLEUCalculator()
score = bleu.bleu("I am waiting on the beach", "He is walking on the beach")
bleu_ja = BLEUCalculator(lang="en")