def __init__(self, lang, use_stemmer, stop_words):
self.lang = lang
self.use_stemmer = use_stemmer
self.word_tokenizer = WordTokenizer(locale=lang)
# As we have only one language currently, no need to
# check if supported
self.stemmer = Stemmer.Stemmer(self.lang)
self.stop_words = stop_words
class Preprocess:
def __init__(self, lang, use_stemmer, stop_words):
self.lang = lang
self.use_stemmer = use_stemmer
self.word_tokenizer = WordTokenizer(locale=lang)
# As we have only one language currently, no need to
# check if supported
self.stemmer = Stemmer.Stemmer(self.lang)
self.stop_words = stop_words
def stem(self, tokens):
return (self.stemmer.stemWords(tokens)
if self.use_stemmer
else tokens)
def remove_accents(self, text):
return (unicodedata.normalize('NFD', text)
.encode('ascii', 'ignore')
.decode("utf-8"))
def lower(self, text):
return text.lower()
def clear_html(self, text):
# Html cleaner will throw exception when an incorrect tag
# formation is detected a 'prompt' like symbol e.g.: '<- '
try:
text = clean_html(html.fromstring(text)).text_content()
except Exception as e:
pass
return text
def is_token_stopword(self, token):
return token in self.stop_words
def is_alpha(self, token):
return token.isalpha()
def filtertokens(self, tokens):
return filter(
lambda t: self.is_alpha(t) and not (self.is_token_stopword(t)),
tokens)
def transform2words(self, text):
return self.word_tokenizer.transform(Sequence(text)).tokens()
def process(self, text):
fnlist = [
self.remove_accents,
self.lower,
self.clear_html,
self.transform2words,
self.filtertokens,
self.stem
]
return compose(*reversed(fnlist))(text)
def main(input_fname, output_fname, lang, to_lower=True):
en_tokenizer = WordTokenizer(locale='en')
fr_tokenizer = WordTokenizer(locale=lang)
def tokenizer(text, tokenizer_fn):
seq = Sequence(text.strip())
return filter(lambda w: w != ' ', tokenizer_fn.transform(seq))
logging.info((lang, "counting pairs"))
counter = Counter()
for line_no, line in enumerate(io.open(input_fname, 'r',
encoding='utf-8')):
if to_lower:
line = line.lower()
parts = line.rstrip().split(' ||| ')
if len(parts) != 4: continue
source_lang, source_text, target_text, count = parts
source_tokens = tokenizer(source_text, en_tokenizer)
target_tokens = tokenizer(target_text, fr_tokenizer)
if len(source_tokens) > 3 or len(target_tokens) > 3: continue
count = int(count)
if count > 1:
if (re.sub('\p{P}', '', source_text[0]) == ''
or re.sub('\p{P}', '', target_text[0]) == ''
or re.sub('\p{P}', '', source_text[-1]) == ''
or re.sub('\p{P}', '', target_text[-1]) == ''):
continue
pair = ' ||| '.join([
source_lang, ' '.join(source_tokens), ' '.join(target_tokens)
])
counter[pair] += count
if line_no % 100000 == 0:
logging.info((lang, line_no))
logging.info((lang, "writing pairs to {0}".format(output_fname)))
with io.open(output_fname, 'w', encoding='utf-8') as out:
for pair, count in counter.most_common():
if count < 10:
break
out.write('{0} ||| {1}\n'.format(pair, count))
def segment(args):
lang = args.lang
w_tokenizer = WordTokenizer(locale=lang)
s_tokenizer = SentenceTokenizer(locale=lang)
if args.only_sent:
for l in args.input:
seq = Sequence(l)
if not seq.empty(): _print(u'\n'.join(s_tokenizer.transform(seq)))
elif args.only_word:
for l in args.input:
seq = Sequence(l)
if not seq.empty(): _print(u' '.join(w_tokenizer.transform(seq)))
else:
for l in args.input:
seq = Sequence(l)
sents = s_tokenizer.transform(seq)
words = w_tokenizer.transform(seq)
for tokenized_sent in words.split(sents):
if not tokenized_sent.empty():
_print(u' '.join(tokenized_sent.tokens()))
def segment(args):
lang = args.lang
w_tokenizer = WordTokenizer(locale=lang)
s_tokenizer = SentenceTokenizer(locale=lang)
if args.only_sent:
for l in args.input:
seq = Sequence(l)
if not seq.empty(): _print(s_tokenizer.transform(seq))
elif args.only_word:
for l in args.input:
seq = Sequence(l)
if not seq.empty(): _print(w_tokenizer.transform(seq))
else:
for l in args.input:
seq = Sequence(l)
sents = s_tokenizer.transform(seq)
words = w_tokenizer.transform(seq)
for tokenized_sent in words.split(sents):
if not tokenized_sent.empty():
_print(u' '.join(tokenized_sent.tokens()))
def main(input_fname, lang, to_lower=True):
en_tokenizer = WordTokenizer(locale='en')
fr_tokenizer = WordTokenizer(locale=lang)
for line_no, line in enumerate(smart_open(input_fname)):
data = json.loads(line)
en_text, pairs = data[0][0][1], data[5]
if not isinstance(pairs, list):
logging.error((input_fname, 'not list', pairs))
continue
for source, _, targets, _, _, _, _ in pairs:
if not isinstance(targets, list):
logging.error((input_fname, 'not list', targets))
continue
for target in targets:
if source in en_text:
count = int(target[1]) or 1
source = tokenizer(source, en_tokenizer, to_lower)
target = tokenizer(target[0], fr_tokenizer, to_lower)
if source and target:
print('{0} ||| {1} ||| {2}'.format(
source, target, count).encode('utf-8'))
if line_no % 10000 == 0:
logging.info((input_fname, line_no))
def word_tokenizer(self):
word_tokenizer = WordTokenizer(locale=self.language.code)
return word_tokenizer
