class NLTKTagger:
'''
class that supplies part of speech tags using NLTK
note: avoids the NLTK downloader (see __init__ method)
'''
def __init__(self):
import nltk
from nltk.tag import PerceptronTagger
from nltk.tokenize import TreebankWordTokenizer
#return pkgutil.get_data('scattertext',
# 'data/viz/semiotic_new.html').decode('utf-8')
path = os.path.dirname(sys.modules['scattertext'].__file__) + '/data/'
tokenizer_fn = path + 'punkt.english.pickle'
tagger_fn = path + 'averaged_perceptron_tagger.pickle'
#tokenizer_fn = os.path.abspath(resource_filename('scattertext.data', 'punkt.english.pickle'))
#tagger_fn = os.path.abspath(resource_filename('scattertext.data', 'averaged_perceptron_tagger.pickle'))
# Load the tagger
self.tagger = PerceptronTagger(load=False)
self.tagger.load(tagger_fn)
# note: nltk.word_tokenize calls the TreebankWordTokenizer, but uses the downloader.
# Calling the TreebankWordTokenizer like this allows skipping the downloader.
# It seems the TreebankWordTokenizer uses PTB tokenization = regexes. i.e. no downloads
# https://github.com/nltk/nltk/blob/develop/nltk/tokenize/treebank.py#L25
self.tokenize = TreebankWordTokenizer().tokenize
self.sent_detector = nltk.data.load(tokenizer_fn)
# http://www.nltk.org/book/ch05.html
def tag_text(self, text):
'''take input text and return tokens w/ part of speech tags using NLTK'''
# putting import here instead of top of file b.c. not all will have nltk installed
sents = self.sent_detector.tokenize(
text
) # TODO: this will fail on some unicode chars. I think assumes ascii
word_pos_pairs = []
all_tokens = []
for sent in sents:
tokens = self.tokenize(sent)
all_tokens = all_tokens + tokens
word_pos_pairs = word_pos_pairs + self.tagger.tag(tokens)
return {
'tokens': all_tokens,
'pos': [tag for (w, tag) in word_pos_pairs]
}
def tag_tokens(self, tokens):
word_pos_pairs = self.tagger.tag(tokens)
return {'tokens': tokens, 'pos': [tag for (w, tag) in word_pos_pairs]}
def smaller_subtree_containing_the_drugs(sentence, target_drugs):
tree_string = nlp.annotate(sentence,
properties={
'annotators': 'parse',
'outputFormat': 'json'
})
tagger = PerceptronTagger()
best_subtree = None
size = 9999999
target_drugs = [dr for drug in target_drugs for dr in drug.split(' ')]
for s in tree_string['sentences']:
tree_parsed = Tree.fromstring(s['parse'])
for subtree in tree_parsed.subtrees():
# print(subtree.pretty_print())
leafs = subtree.leaves()
current_size = len(leafs)
if all_drugs_in_tree(target_drugs, leafs):
if current_size < size:
best_subtree = subtree
size = current_size
# print(subtree.leaves())
try:
clean = clean_sentence(best_subtree.leaves())
except:
clean = clean_sentence(sentence.split())
# print('clean',clean)
tagged = tagger.tag(clean)
# print('tag:', tagged)
lemmatized = preprocessor_lemmatize(tagged)
# print('lemmatized', lemmatized)
new_sentence = ' '.join([l for l, t in lemmatized])
return new_sentence
def count_common_propper_nouns(self, s0, s1):
tagger = PerceptronTagger()
s0_tags = tagger.tag(s0)
s1_tags = tagger.tag(s1)
NNP_s0 = [values[0] for values in s0_tags if values[1] == 'NNP']
NNP_s1 = [values[0] for values in s1_tags if values[1] == 'NNP']
return len(set(NNP_s0) & set(NNP_s1))
def common_description(self, s0, s1):
tagger = PerceptronTagger()
s0_tags = tagger.tag(s0)
s1_tags = tagger.tag(s1)
total_dist = 0
for word, tag in s0_tags:
if tag.startswith('N') or tag.startswith('V') or tag.startswith('J') or tag.startswith('R'):
max_dist = 0
for synset in wn.synsets(word, self.penn_to_wn(tag)):
desc = word_tokenize(synset.definition())
dist = len(list(set(s1) & set(desc)))
if dist > max_dist:
max_dist = dist
total_dist += max_dist
for word, tag in s1_tags:
if tag.startswith('N') or tag.startswith('V') or tag.startswith('J') or tag.startswith('R'):
max_dist = 0
for synset in wn.synsets(word, self.penn_to_wn(tag)):
desc = word_tokenize(synset.definition())
dist = len(list(set(s0) & set(desc)))
if dist > max_dist:
max_dist = dist
total_dist += max_dist
return total_dist
class CountAdjectives(TransformerMixin):
""" adds postags, learns weights """
def __init__(self):
super(CountAdjectives, self).__init__()
self.tagger = PerceptronTagger(load=True)
training_corpus = list(alpino.tagged_sents())
self.tagger.train(training_corpus)
def postag(self, x):
postagged = self.tagger.tag(x.split())
onlytags = [tt[1] for tt in postagged]
return onlytags
def count_adjectives(self, x):
postagged = self.postag(x)
totalcount = len(postagged)
adjlength = postagged.count('adj')
if adjlength > 0:
return adjlength / totalcount
return 0
def transform(self, X, y=None):
new_X = [[self.count_adjectives(x)] for x in X]
return new_X
def fit(self, X, y=None):
return self
class NLTKTagger:
'''
class that supplies part of speech tags using NLTK
note: avoids the NLTK downloader (see __init__ method)
'''
def __init__(self):
import nltk
from nltk.tag import PerceptronTagger
from nltk.tokenize import TreebankWordTokenizer
#return pkgutil.get_data('scattertext',
# 'data/viz/semiotic_new.html').decode('utf-8')
path = os.path.dirname(sys.modules['scattertext'].__file__)+'/data/'
tokenizer_fn = path + 'punkt.english.pickle'
tagger_fn = path + 'averaged_perceptron_tagger.pickle'
#tokenizer_fn = os.path.abspath(resource_filename('scattertext.data', 'punkt.english.pickle'))
#tagger_fn = os.path.abspath(resource_filename('scattertext.data', 'averaged_perceptron_tagger.pickle'))
# Load the tagger
self.tagger = PerceptronTagger(load=False)
self.tagger.load(tagger_fn)
# note: nltk.word_tokenize calls the TreebankWordTokenizer, but uses the downloader.
# Calling the TreebankWordTokenizer like this allows skipping the downloader.
# It seems the TreebankWordTokenizer uses PTB tokenization = regexes. i.e. no downloads
# https://github.com/nltk/nltk/blob/develop/nltk/tokenize/treebank.py#L25
self.tokenize = TreebankWordTokenizer().tokenize
self.sent_detector = nltk.data.load(tokenizer_fn)
# http://www.nltk.org/book/ch05.html
def tag_text(self, text):
'''take input text and return tokens w/ part of speech tags using NLTK'''
# putting import here instead of top of file b.c. not all will have nltk installed
sents = self.sent_detector.tokenize(text) # TODO: this will fail on some unicode chars. I think assumes ascii
word_pos_pairs = []
all_tokens = []
for sent in sents:
tokens = self.tokenize(sent)
all_tokens = all_tokens + tokens
word_pos_pairs = word_pos_pairs + self.tagger.tag(tokens)
return {'tokens': all_tokens, 'pos': [tag for (w, tag) in word_pos_pairs]}
def tag_tokens(self, tokens):
word_pos_pairs = self.tagger.tag(tokens)
return {'tokens': tokens, 'pos': [tag for (w, tag) in word_pos_pairs]}
def tagger(self):
"""
Usage:
training_corpus = list(alp.tagged_sents())
tagger = PerceptronTagger(load=True)
tagger.train(training_corpus)
#sent = 'NLTK is een goeda taal voor het leren over NLP'.split()
print(tagger.tag(article_text.split()))
:return:
"""
# Load Corpus
training_corpus = list(alp.tagged_sents())
tagger = PerceptronTagger(load=True)
# Build tagger
tagger.train(training_corpus)
return tagger.tag(self.string.split())
token.shape_, token.is_alpha, token.is_stop,[child for child in token.children]]
if dependency[0] == "\n":
whole_sen.append(parsed)
parsed=[]
else:
parsed.append(dependency)
frysian=[]
tagger = PerceptronTagger()
with open('frysian_data.txt', 'r', encoding="utf-8") as fr_infile:
for sentence in fr_infile:
sentence = word_tokenize(sentence)
pos = tagger.tag(sentence)
fr.append(pos)
other=[]
final =[]
fr_longer=[]
for k in range(len(fr)):
fries = fr[k]
parsed = whole_sen[k]
if len(fries) == len(parsed):
for words, fr_words in zip(parsed,fries):
print(words[0])
other.append(words)
# print(fr_words[0])
words[0]=fr_words
final.append(words)
def main(file_input):
data_df = pd.read_csv(str(file_input) + '.csv')
data_df = shuffle(data_df)
print("Loaded .csv file Successfully")
print("Total Number of Samples:", data_df.shape[0])
print("Total Number of Features:", data_df.shape[1])
# Missing Values
# column with maximum missing values
def missing_value(data_df):
while data_df.isnull().sum().values.sum() != 0:
col_with_missing_val = (data_df.isnull().sum()).argmax()
data_df = data_df[data_df[col_with_missing_val].notnull(
)] # drop corresponding rows that has NaN values
print("Missing Values in Features:", col_with_missing_val)
return data_df
# Missing Value Treatment:
print("Missing Value Treatment : Start")
data_df = missing_value(data_df)
print("Missing Value Treatment : Stop")
print("Total Number of Samples:", data_df.shape[0])
print("Total Number of Features:", data_df.shape[1])
# pattern matcher for candidate feature
# newly Added Features : Dates format, currency format, number of digits per candidate, number of separators
# per candidate
print("Computing Pattern Transformers: Start")
pattern_strictlyDigits = "^[0-9]*$"
pattern_endWithCharacters = "^\d*[\/.,@$!)(]$" # Only digits + end with special characters
pattern_telephone = "^0[0-9]{12}$"
pattern_vat = "^0?[0-9]{9}$"
pattern_date = '^[0-3]?[0-9](\/|\,|\.|\-){1}[0-9]?[0-9](\/|\,|\.|\-){1}[0-2][0-9]{1,3}$'
pattern_currency_1 = '^[0-9]\.[0-9]+\,[0-9]*$' # captures ddddd,dddd
pattern_currency_2 = '^[0-9]+\,[0-9]+$'
data_df['currency_filter'] = data_df['candidate'].str.contains(pattern_currency_1, regex=True).astype(np.int64)\
| data_df['candidate'].str.contains(pattern_currency_2, regex=True).astype(np.int64)
data_df['dates_filter'] = data_df['candidate'].str.contains(
pattern_date, regex=True).astype(np.int64)
data_df["Is_strictly_Digits"] = data_df["candidate"].str.contains(
pattern_strictlyDigits, regex=True).astype(np.int64)
data_df["endWithCharacters"] = data_df["candidate"].str.contains(
pattern_endWithCharacters, regex=True).astype(np.int64)
data_df["Number_of_Digits"] = data_df['candidate'].apply(
lambda x: len(re.sub("\W", "", x)))
data_df["Number_of_Separators"] = data_df['candidate'].apply(
lambda x: len(re.sub("\w", "", x)))
data_df["Length_of_Candidate"] = data_df['candidate'].apply(
lambda x: len(x))
# included the country code
data_df["Telephone"] = data_df["candidate"].str.contains(
pattern_telephone, regex=True).astype(np.int64)
# VAT number contains 9 to 10 digits
data_df["VATNumber"] = data_df["candidate"].str.contains(
pattern_vat, regex=True).astype(np.int64)
# drop blacklisted variables
dates_index = data_df.index[data_df['dates_filter'] == 1].tolist()
data_df = data_df.drop(index=dates_index, axis=0)
data_df = data_df.drop("dates_filter", axis=1)
currency_index = data_df.index[data_df['currency_filter'] == 1].tolist()
data_df = data_df.drop(index=currency_index, axis=0)
data_df = data_df.drop(["currency_filter"], axis=1)
telephone_index = data_df.index[data_df['Telephone'] == 1].tolist()
data_df = data_df.drop(index=telephone_index, axis=0)
data_df = data_df.drop(["Telephone"], axis=1)
vat_index = data_df.index[data_df['VATNumber'] == 1].tolist()
data_df = data_df.drop(index=vat_index, axis=0)
data_df = data_df.drop(["VATNumber"], axis=1)
vat_index = data_df.index[data_df['endWithCharacters'] == 1].tolist()
data_df = data_df.drop(index=vat_index, axis=0)
data_df = data_df.drop(["endWithCharacters"], axis=1)
print("Computing Pattern Transformers: Stop")
# NLP Techniques:
# Tokenization, Stemming, lemmatization, Frequency Distribution, Bag of words approach
# Combine three text columns to single column - This columns contains he full text
data_df["Text"] = data_df["line_before"] + data_df["line_at"] + data_df[
"line_after"]
print("Computing Context Transformers: Start")
# Context Transformers
def email_match(doc):
match = re.search(r'[\w\.-]+@[\w\.-]+', str(doc))
if match != None:
return 1
else:
return 0
data_df["Number_of_Characters_Text"] = data_df["Text"].apply(
lambda x: len(re.sub("[^a-z]", "", str(x))))
data_df["Number_of_Digits_Text"] = data_df["Text"].apply(
lambda x: len(re.sub("[^0-9]+", "", str(x))))
data_df["Number_of_Separators_Text"] = data_df["Text"].apply(lambda x: len(
(re.sub("[\w]+", "", str(x))).replace(" ", "")))
data_df["Is_Email_Exists"] = data_df["Text"].apply(
email_match) # place 1 everywhere email found else 0
data_df["Number_of_spaces"] = data_df["Text"].apply(
lambda x: str(x).count(' ')) # counts number of spaces,
# Clean Data - Tokenization, Stop word check, Size filter, Stemming - Dutch Language
ss = SnowballStemmer("dutch", "french")
def clean_data(doc):
ignore = list(set(stopwords.words(
'dutch', 'french'))) # ignore the list of stopwords
exl_chars = list(set(string.punctuation))
exl_chars.append('€')
# remove email ids to avoid conflicts in vocabulary construction
doc = re.sub("[\w\.-]+@[\w\.-]+", " ", str(doc))
doc = re.sub("\d", " ", str(doc))
doc = ''.join([ch for ch in doc if ch not in exl_chars])
words = []
for i in word_tokenize(doc): # tokenization
if i not in ignore:
if len(i) >= 2: # standalone letters do not add any value
i = ss.stem(i)
words.append(i)
doc = ' '.join(list(set(words)))
return doc
print("Cleaning Text Data: Start")
data_df["Text"] = data_df["Text"].apply(
clean_data) # tokenize, stem and lammetize
print("Cleaning Text Data: Stop")
print("Computing POS Vectors: Start")
# training_corpus = alp.tagged_sents()
alp_tagged_sent = list(alp.tagged_sents())
tagger = PerceptronTagger(load=False)
tagger.train(alp_tagged_sent)
def count_adj(doc):
tags = tagger.tag(doc.split())
for tup in tags:
first_3_characters = tup[0][:3]
last_3_characters = tup[0][3:]
if len(tags[0]) >= 3 and first_3_characters[
0] == first_3_characters[1] == first_3_characters[2]:
tags.remove(tup)
if len(tags[0]) >= 3 and last_3_characters[0] == last_3_characters[
1] == last_3_characters[2]:
tags.remove(tup)
counts = Counter(tag for word, tag in tags)
count_adj_adv = counts['adv'] + counts['adj']
return count_adj_adv
def count_nn(doc):
tags = tagger.tag(doc.split())
for tup in tags:
first_3_characters = tup[0][:3]
last_3_characters = tup[0][3:]
if len(tags[0]) >= 3 and first_3_characters[
0] == first_3_characters[1] == first_3_characters[2]:
tags.remove(tup)
if len(tags[0]) >= 3 and last_3_characters[0] == last_3_characters[
1] == last_3_characters[2]:
tags.remove(tup)
counts = Counter(tag for word, tag in tags)
count_nn = counts['noun']
return count_nn
def count_verb(doc):
tags = tagger.tag(doc.split())
for tup in tags:
first_3_characters = tup[0][:3]
last_3_characters = tup[0][3:]
if len(tags[0]) >= 3 and first_3_characters[
0] == first_3_characters[1] == first_3_characters[2]:
tags.remove(tup)
if len(tags[0]) >= 3 and last_3_characters[0] == last_3_characters[
1] == last_3_characters[2]:
tags.remove(tup)
counts = Counter(tag for word, tag in tags)
count_verb = counts['verb']
return count_verb
data_df["Adv_Adj_Count"] = data_df["Text"].apply(count_adj)
data_df["NN_count"] = data_df["Text"].apply(count_nn)
data_df["Verb_count"] = data_df["Text"].apply(count_verb)
print("Computing POS Vectors: Stop")
print("Computing Vocabulary: Start")
# store all the words in positive class and negative in two separate lists
docs_pos = []
docs_pos.extend(
word_tokenize(words) for words in data_df.Text[data_df.gold == 1])
docs_pos = list(itertools.chain(*docs_pos))
# Clean text data - remove words like --- iiiiiii, hhhhhccchhhh, abvwwwwwcgdccc
for i in docs_pos:
first_3_characters = i[:3]
last_3_characters = i[-3:]
if len(i) >= 3 and first_3_characters[0] == first_3_characters[
1] == first_3_characters[2]:
docs_pos.remove(i)
if i in docs_pos and len(i) >= 3 and last_3_characters[
0] == last_3_characters[1] == last_3_characters[2]:
docs_pos.remove(i)
print("Positve class words are stored successfully")
all_words_pos = nltk.FreqDist(docs_pos)
print("Computing vocabulary based on Positive Class")
# find popular words, popular equals more than 25 times in the corpus
popular_pos_words = []
for i in all_words_pos.items():
if i[1] >= 25:
popular_pos_words.append(i[0])
# Filter nouns from the popular positive class words
tagged_pos_words = tagger.tag(popular_pos_words)
filtered_tag_pos_words_nouns = []
for word in tagged_pos_words:
if word[1] == 'noun':
filtered_tag_pos_words_nouns.append(word[0])
vocab_pos = list(set(filtered_tag_pos_words_nouns))
vocabulary = list(set(vocab_pos))
# save vocabulary
with open("vocab.txt", "wb") as fp:
pickle.dump(vocabulary, fp)
print("Computing Vocabulary: Stop")
print("Length of Vocabulary: ", len(vocabulary))
print("Computing Bag of Words Vectors: Start")
def build_features(doc):
vector = np.zeros((1, len(vocabulary)), dtype=np.int64)
for w in word_tokenize(doc):
for idx, vocab in enumerate(vocabulary):
if vocab == w:
vector[0][idx] += 1
return vector
bag_vectors = data_df["Text"].apply(build_features)
feature_vectors = np.zeros((data_df.shape[0], len(vocabulary)),
dtype=np.int64)
for pos, index in enumerate(data_df.index.values):
feature_vectors[pos, :] = bag_vectors[index]
cols = ["BOW_" + str(col) for col in range(0, len(vocabulary))]
for col_index, col in enumerate(cols):
data_df[col] = feature_vectors[:,
col_index].reshape(data_df.shape[0], 1)
print("Computing Bag of Words Vectors: Stop")
print("Computing Context Transformers: Stop")
print("Computing Location Transformers: Start")
data_df["location_page_nr"] = data_df["page_nr"].apply(lambda x: 100
if x >= 50 else x)
data_df["location_line_nr"] = data_df["line_nr"].apply(lambda x: 100
if x >= 50 else x)
print("Computing Location Transformers: Stop")
print("Total Number of Newly Added Features:", data_df.shape[1] - 7)
print("Building ML - Neural Network Model: Start")
X = data_df.drop([
"candidate", "Text", "gold", "label", "line_after", "line_at",
"line_before", "line_nr", "page_nr"
],
axis=1)
y = data_df.gold
# Normalisation
X = (X - X.mean(axis=0)) / X.std(axis=0)
def build_model(input_shape):
model = Sequential()
model.add(Dense(1024, input_shape=(input_shape, )))
model.add(Activation('sigmoid'))
model.add(Dense(512))
model.add(Activation('sigmoid'))
model.add(Dense(128))
model.add(Activation('sigmoid'))
model.add(Dense(1, activation="sigmoid"))
model.compile(optimizer='adam',
loss=tf.keras.losses.mean_squared_error,
metrics=['accuracy'])
return model
# Stratified k-Fold
k_fold_outer = model_selection.StratifiedKFold(n_splits=5)
scores = []
split = 0
for train_index, test_index in k_fold_outer.split(X, y):
X_train, X_val = X.iloc[train_index], X.iloc[test_index]
y_train, y_val = y.iloc[train_index], y.iloc[test_index]
model = build_model(X_train.shape[1])
history = model.fit(X_train,
y_train,
epochs=5,
batch_size=1024,
verbose=1)
results = model.evaluate(X_val, y_val)
scores.append(results[1])
split += 1
del model, history, results
model = build_model(X.shape[1])
model.fit(X, y, verbose=0)
print('Saving the Model *.h5...')
model.save('model_candidate_filter.h5')
yHat_proba = model.predict(X)
yHat = np.copy(yHat_proba)
yHat[yHat <= 0.5] = 0
yHat[yHat > 0.5] = 1
br_score = np.around(metrics.brier_score_loss(y, yHat_proba, pos_label=1),
decimals=5)
print("Storing Results in .csv file")
confidence = np.zeros((yHat_proba.shape[0], yHat_proba.shape[1]))
for i in range(0, yHat_proba.shape[0]):
if yHat_proba[i] <= 0.5:
confidence[i] = 1 - yHat_proba[i]
else:
confidence[i] = yHat_proba[i]
results_data_frame = pd.DataFrame(
columns=["Predictions", "Confidence Level"], index=data_df.index)
results_data_frame["Predictions"] = yHat.astype(np.int64).ravel()
results_data_frame["Confidence Level"] = np.around(confidence, decimals=4)
results_data_frame.to_csv("Results_predictions_confidence_train.csv",
encoding='utf-8',
header=True,
index=True)
return np.mean(scores), br_score
def count_verbs(self, s0):
tagger = PerceptronTagger()
s0_tags = tagger.tag(s0)
V_s0 = [values[0] for values in s0_tags if values[1] == 'VBP']
return len(V_s0)
def count_nouns(self, s0):
tagger = PerceptronTagger()
s0_tags = tagger.tag(s0)
NN_s0 = [values[0] for values in s0_tags if values[1] == 'NN']
return len(NN_s0)
noga = len(count_good_actors)
#number of good actors
nogc = len(count_good_actions)
PICKLE = "taggers/averaged_perceptron_tagger/averaged_perceptron_tagger.pickle"
import nltk.data
from nltk.tag import PerceptronTagger
_nltk_pos_tagger = PerceptronTagger(load=False)
_nltk_pos_tagger.load(PICKLE)
print(count_good_actors)
S = np.zeros(shape=(nos, noga + nogc))
i = 0
for sent_pos in tokenized_actors:
for token1 in sent_pos:
j = 0
tt1 = _nltk_pos_tagger.tag([token1])
for feature in count_good_actors:
ft = _nltk_pos_tagger.tag([feature])
simval = word_sim(tt1[0], ft[0], i)
S[i][j] = S[i][j] + simval
j = j + 1
i = i + 1
i = 0
for sent_pos in tokenized_actions:
for token1 in sent_pos:
j = noga
tt1 = _nltk_pos_tagger.tag([token1])
for feature in count_good_actions:
ft = _nltk_pos_tagger.tag([feature])
simval = word_sim(tt1[0], ft[0], i)
for sent1, sent2 in zip(predicted_text, tagged_text):
scores.append(score(sent1, sent2))
average_score = sum(scores) / len(predicted_text)
return average_score
# Tagged text remains the same from the Penn Treebank
tagged_text = list(treebank.tagged_sents())
# 1. Perceptron Tagger
from nltk.tag import PerceptronTagger
data = list(data)
tagger = PerceptronTagger()
predicted_text = []
for sent in data:
predicted_text.append(tagger.tag(sent))
perceptron_score = score_dataset(predicted_text, tagged_text)
# 1.804
# 2. nltk.Pos_tag Tagger
from nltk import pos_tag
predicted_text = []
for sent in data:
predicted_text.append(pos_tag(sent))
nltk_score = score_dataset(predicted_text, tagged_text)
# 1.804
# 3. Spacy POS Tagger
import spacy
