def txt_transformer(file_conllu):
s_list = list()
with open(file_conllu, 'r') as f:
ok = f.read()
try:
conll = pc.load_from_string(ok)
except pc.exception.ParseError:
conll = pc.load_from_string(to_conllu(ok))
for s in conll:
s_txt = ""
for word in s[:-1]:
s_txt = s_txt + " " + word.form
s_txt = s_txt.strip() + ".\n"
s_list.append(s_txt)
return u''.join(s_list).encode('utf-8')
def load_sentences(n, f):
count = 0
lines = ""
line = f.readline()
while line and count < n:
lines += line # line.replace("sent_id", "sent_id =") if (opt.use_v1 and opt.get_ids) else line
if line == "\n":
count += 1
line = f.readline()
not_empty = True if line else False
try:
conll = load_from_string(lines)
except Exception:
conll = load_from_string("")
print("bad conll")
return conll, not_empty
def load_predictions(args):
# Regular CoNLL-U format
if args.pred_upos_index == 3 and args.pred_xpos_index == 4 and args.pred_feats_index == 5:
return pyconll.load_from_file(args.prediction)
# other format
else:
s = ""
with open(args.prediction, 'r') as pred_file:
for line in pred_file:
if line.strip() == "":
s += line
elif line.startswith("#"):
s += line
else:
elements = line.split("\t")
if args.pred_upos_index >= 0 and args.pred_upos_index < len(elements):
upos = elements[args.pred_upos_index].strip()
else:
upos = "_"
if args.pred_xpos_index >= 0 and args.pred_xpos_index < len(elements):
xpos = elements[args.pred_xpos_index].strip()
else:
xpos = "_"
if args.pred_feats_index >= 0 and args.pred_feats_index < len(elements):
feats = elements[args.pred_feats_index].strip()
else:
feats = "_"
s += "0\t_\t_\t{}\t{}\t{}\t0\t_\t_\t_\n".format(upos, xpos, feats)
return pyconll.load_from_string(s)
def __call__(self, text):
sentences = self.tokenize(text)
for s in sentences:
self.tag(s)
self.parse(s)
conllu = self.write(sentences, "conllu")
result = pyconll.load_from_string(conllu)
return result
def parse_syntax(self, text_str):
processed = self.udpipe_pipeline.process(text_str, self.udpipe_error)
if self.udpipe_error.occurred():
logging.error("An error occurred when running run_udpipe: %s", self.udpipe_error.message)
return None
parsed_data = pyconll.load_from_string(processed)[0]
return parsed_data
def __call__(self, text, ner_list):
sentences = self.tokenize(text)
for s in sentences:
self.tag(s)
self.parse(s)
conllu = self.write(sentences, "conllu")
result = pyconll.load_from_string(conllu)
for word, ner_tag in zip(result[0], ner_list):
word.misc = ner_tag
return result
def test_load_from_string():
"""
Test that a CoNLL file can properly be loaded from a string.
"""
with open(fixture_location('basic.conll')) as f:
contents = f.read()
c = load_from_string(contents)
sent = c[1]
assert len(c) == 4
assert len(sent) == 14
assert sent['10'].form == 'donc'
def add_speakers(conll_string, path_to_csv):
import pyconll
import pandas as pd
df = pd.read_csv(path_to_csv)
speaker_list = df["Speaker"]
conll = pyconll.load_from_string(conll_string)
assert len(conll) == len(speaker_list), \
"List of speakers and list of sentences not in sync ({} speakers, {} sentences)".format(len(speaker_list),
len(conll))
for c, s in zip(conll, speaker_list):
c.set_meta("speaker", s)
return conll
def get_triples(input_part=9100):
"""
Empty nsubj and dobj are represented by empty string.
"""
config = configparser.ConfigParser()
config.read('config.ini')
outfilen = os.path.join(config['DEFAULT']['ProjectDirectory'],
f'dataframe/freq{input_part}.pkl')
if os.path.exists(outfilen):
logging.info(f'File exists: {outfilen}')
return
triples = []
for filen in sorted(
glob(
f'/mnt/permanent/Language/English/Crawl/DepCC/corpus/parsed/part-m-*{input_part}.gz'
)):
logging.info(filen)
for sent_str in depcc_to_conllu(filen):
try:
train = pyconll.load_from_string(sent_str)
except Exception as e:
logging.error(e)
continue
sentence = train.pop() # sent_str is only one sentence.
triples_in_sent = defaultdict(lambda: {'nsubj': '', 'dobj': ''})
# triples = {id_of_root: {'nsubj': 'dog'}}
# Collecting the arguments..
for token in sentence:
if token.deprel in ['nsubj', 'dobj']:
triples_in_sent[token.head][token.deprel] = token.lemma
# Collecting the verbs, not only the main pred of the sentence.
for id_form_1 in triples_in_sent:
if sentence[int(id_form_1)].upos.startswith('V'):
verb = sentence[int(id_form_1)].lemma
triples_in_sent[id_form_1]['ROOT'] = verb
# Appending full triples to the list..
for triple in triples_in_sent.values():
if 'ROOT' in triple:
triples.append(triple)
df = pd.DataFrame(triples)
df = df.groupby(list(df.columns)).size().sort_values(ascending=False)
df.to_frame(name='freq').to_pickle(outfilen)
def test_load_from_file_and_string_equivalence():
"""
Test that the Conll object created from a string and file is the same if
the underlying source is the same.
"""
with open(fixture_location('long.conll')) as f:
contents = f.read()
str_c = load_from_string(contents)
file_c = load_from_file(fixture_location('long.conll'))
assert len(str_c) == len(file_c)
for i in range(len(str_c)):
assert str_c[i].id == file_c[i].id
assert str_c[i].text == file_c[i].text
print(str_c[i].conll())
print(file_c[i].conll())
for str_token in str_c[i]:
file_token = file_c[i][str_token.id]
assert_token_members(str_token, file_token.id, file_token.form,
file_token.lemma, file_token.upos,
file_token.xpos, file_token.feats,
file_token.head, file_token.deprel,
file_token.deps, file_token.misc)
def check(self, sentence):
doc_de = self.nlp_de(sentence)
conllu = doc_de.conll_file.conll_as_string()
conll = pyconll.load_from_string(conllu)
print(dir(conll))
translate_client = translate.Client()
english = translate_client.translate(
sentence, target_language='en')['translatedText']
print("ENGLISH " + html.unescape(english))
doc_en = self.nlp_en(html.unescape(english))
#print(doc_de.conll_file.conll_as_string())
sentences_ret = []
for sentence in doc_de.sentences:
words_ret = [ReturnWord() for _ in range(len(sentence.words))]
for word in sentence.words:
#print("[%d] Word: %s" % (int(word.index), word.text))
eng = self.translator.translate(word.lemma,
src='de',
dest='en')
#print(dir(words_ret[int(word.index) - 1].translations))
#print(format(eng.extra_data))
#print(format(eng.extra_data['all-translations']))
words_ret[int(word.index) - 1].translation = eng.text
if (eng.extra_data['all-translations'] is not None):
words_ret[int(word.index) - 1].other_translations.extend(
eng.extra_data['all-translations'][0][1])
words_ret[int(word.index) - 1].set_vars(
word.text, int(word.index), word.upos, word.xpos,
word.governor, word.dependency_relation)
#print("%s, %s, %s, %s" % (words_ret[int(word.index) - 1].text, words_ret[int(word.index) - 1].upos, words_ret[int(word.index) - 1].xpos, words_ret[int(word.index) - 1].relation))
if (word.upos == 'ADP'):
#print("%s is ADP" % word.text)
if (word.lemma in self.prep_akk):
#print("%s is accusative" % word.text)
words_ret[int(word.index) - 1].notes.append(
word.text + " takes the accusative case.")
words_ret[word.governor - 1].notes.append(
"Because " + word.text + " is accusative, so is " +
sentence.words[word.governor - 1].text)
words_ret[word.governor - 1].case = "Acc"
elif (word.lemma in self.prep_dat):
#print("%s is dative" % word.text)
words_ret[int(word.index) -
1].notes.append(word.text +
" takes the dative case.")
words_ret[word.governor - 1].notes.append(
"Because " + word.text + " is dative, so is " +
sentence.words[word.governor - 1].text)
words_ret[word.governor - 1].case = "Dat"
elif (word.lemma in self.prep_acc_dat):
#print("%s is akk/dat" % word.text)
words_ret[int(word.index) - 1].notes.append(
word.text +
" takes Dative if it answers the question 'where?'"
)
words_ret[int(word.index) - 1].notes.append(
word.text +
" takes the Accusative if it answers the question 'where to?'"
)
words_ret[word.governor - 1].notes.append(
sentence.words[word.governor - 1].text +
" takes either the Accusative or Dative case.")
words_ret[word.governor - 1].case = "Acc|Dat"
if (word.upos == 'NOUN'):
words_ret[int(word.index) - 1].gender = (self.nlp_de(
word.text).sentences[0].words[0].feats.split('|')
[1].split('=')[1])
#print("%s, %s" % (words_ret[int(word.index) - 1].gender, word.text))
feats = word.feats.split('|')
for feat in feats:
pair = feat.split('=')
#print("%s: %s" % (word.text, pair))
if (pair[0] == "Gender"
and words_ret[int(word.index) - 1].gender == ""):
words_ret[int(word.index) - 1].gender = pair[1]
elif (pair[0] == "Number"):
words_ret[int(word.index) - 1].number = pair[1]
elif (pair[0] == "Case"
and words_ret[int(word.index) - 1].case == ""):
words_ret[int(word.index) - 1].case = pair[1]
#if(len(pair) is 2): print("%s gender: %s (pair[1] is %s)" % (word.text, words_ret[int(word.index) - 1].gender, pair[1]))
if (word.dependency_relation == 'obj'
or word.dependency_relation == 'dobj'):
words_ret[int(word.index) - 1].notes.append(
word.text +
" is probably a direct object, which takes the accusative case."
)
words_ret[int(word.index) - 1].case = "Acc"
if (word.dependency_relation == 'iobj'):
words_ret[int(word.index) - 1].notes.append(
word.text +
" is probably an indirect object, which takes the dative case."
)
words_ret[int(word.index) - 1].case = "Dat"
for word in sentence.words:
#print("%s == 'ART' ? %s" % (word.xpos, (word.xpos == 'ART')))
if (word.xpos == 'ART'):
#print(words_ret[word.governor - 1].case)
if ('Acc' in words_ret[word.governor - 1].case):
#print("%s is Acc %s" % (str(words_ret[word.governor - 1].text), str(words_ret[word.governor - 1].case)))
nom = self.nominative[words_ret[word.governor -
1].gender]
article = self.accusative[words_ret[word.governor -
1].gender]
words_ret[int(word.index) - 1].notes.append(
"If %s %s takes the accusative case the article should be %s"
%
(nom, words_ret[word.governor - 1].text, article))
if ('Dat' in words_ret[word.governor - 1].case):
#print("%s is Dat %s" % (str(words_ret[word.governor - 1].text), str(words_ret[word.governor - 1].case)))
nom = self.nominative[words_ret[word.governor -
1].gender]
article = self.dative[words_ret[word.governor -
1].gender]
words_ret[int(word.index) - 1].notes.append(
"If %s %s takes the dative case the article should be %s"
%
(nom, words_ret[word.governor - 1].text, article))
if ('Gen' in words_ret[word.governor - 1].case):
#print("%s is Gen %s" % (str(words_ret[word.governor - 1].text), str(words_ret[word.governor - 1].case)))
nom = self.nominative[words_ret[word.governor -
1].gender]
article = self.genitive[words_ret[word.governor -
1].gender]
words_ret[int(word.index) - 1].notes.append(
"If %s %s takes the genitive case the article should be %s"
%
(nom, words_ret[word.governor - 1].text, article))
return new_sents, alt_feat_dict
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="generate noisy outputs")
parser.add_argument("-unimorph", type=str, help="unimorph file for lg")
parser.add_argument("-orig", type=str, help="original conllu")
parser.add_argument("-alt", type=str, help="altered conllu")
args = parser.parse_args()
treebank = args.orig
new_treebank = args.alt
data = pyconll.load_from_file(treebank)
um = read_um(args.unimorph)
um2ud, ud2um = read_ud_um()
random.seed(23)
new_sents, alt_feat_dict = sample_noise(data, um, um2ud)
with open(new_treebank, "w") as op:
for new_sent in new_sents:
sent_ = pyconll.load_from_string("\n".join(new_sent))
assert len(sent_) == 1, "more than one sentence"
# op.write("# text = " + " ".join([t_.form for t_ in sent_[0] if t_.form]) + "\n")
op.write("\n".join(new_sent) + "\n\n")
alt_feat_dict = sorted(alt_feat_dict.items(), key=lambda x: x[0])
print(args.unimorph.split("/")[-1])
for (pos, feat), count in alt_feat_dict:
print("%s:%s\t%.2f" % (pos, feat, count * 100 / len(new_sents)))
