def add_line_break(self, text):
# because write to a pseudo file is much faster
# here we use StringIO instead of string concatenation
new_text = StringIO()
# generate features for each word to predict if there is a line break after the word
# split text by blank
features = word_features(text, gold_standard=False).values
for row in features:
feature_string = " ".join([str(i) for i in row])
self.tagger.add(feature_string)
# generate prediction
self.tagger.parse()
# change text according to prediction
for idx in range(0, len(features)):
new_text.write(features[idx][0])
# If the word has new line after it, add new line
# else add whitespace after word
if self.tagger.y2(idx) == "NL":
new_text.write("\n\n")
else:
new_text.write(" ")
# clear parsed words
self.tagger.clear()
# to join string afterwards is more efficient in python
return new_text.getvalue()
def prepare_crf_data(dir, output_file, file_names=[]):
'''
Parse data in .xml and .txt file to data that can be recognized by crf++ tool
Export all newly adapted data to output_file
'''
# Store transferred data, which is used for crf training
total_features = DataFrame()
# Store the begin train time
begin_train = time.time()
print("Data preparation begin:\n")
if len(file_names) > 0:
files = file_names
else:
files = os.listdir(dir)
for file in files:
with codecs.open(dir + file, "r", "ISO-8859-1") as f:
string = f.read()
if file.endswith(".xml"):
text = xml_strip(string)
assert len(text) < len(string)
elif file.endswith(".txt"):
text = string
else: continue
total_features = pd.concat([total_features, word_features(text)], axis=0)
# print total training time
end_train = time.time() - begin_train
print("Done prepping training data in ", end_train, "seconds")
# export training data
total_features.to_csv(output_file, encoding='utf-8', sep="\t", header=False, index=False)
print "Training features saved"
