def transform_20ng_dataset(ouput_file):
"""
Preprocess the 20NG dataset and stores the preprocessed data
:param ouput_file:
:return:
"""
train_data, train_labels = get_20ng(split='train')
test_data, test_labels = get_20ng(split='test')
wv = WordVectorizer()
wv.fit(train_data)
glove_embedding = wv.get_glove_embedding()
train_data = wv.transform(train_data)
test_data = wv.transform(test_data)
data = {}
data['train_data'] = train_data
data['test_data'] = test_data
data['train_labels'] = train_labels
data['test_labels'] = test_labels
data['word2id'] = wv.word2id
data['id2word'] = wv.id2word
data['embedding'] = glove_embedding
with open(ouput_file, "wb") as f:
pickle.dump(data, f, protocol=pickle.HIGHEST_PROTOCOL)
def __init__(self, word2vec_model_path, max_tokens_length, position_vector=True, word_position_size=10, ner=False, pos=False, dependency=False):
self.position_vector = position_vector
# word_position vectors will be filled when calling init_size function
self.word_position = None
self.word_position_size = word_position_size
self.sentence_vectorizer = WordVectorizer(word2vec_model_path, ner=ner, pos=pos, dependency=dependency)
# sizes of the output sequence matrix m is number of words in the sequence
# n is the size of the vector representation of each word in the sequence
self.m = max_tokens_length
self.n = self.sentence_vectorizer.model.vector_size + 2 * self.word_position_size
# original index = -l+1,....,0,...l-1
# array index = 0,.......,(l-1),...(2xl)-1
self.word_position = np.random.rand((2 * max_tokens_length) - 1, self.word_position_size)
import os
import gensim
from tweet import Tweet
from word_vectorizer import WordVectorizer
import csv
#load the word2vec module
word2vecModule = gensim.models.KeyedVectors.load_word2vec_format(
'./embed_tweets_de_200M_200D/embedding_file', binary=False)
#initialize thw WordVectorizer model
word_vectorizer = WordVectorizer(word2vecModule)
# load train data file to get the tweets
with open('../tweetsCrawler/train.csv', 'r') as train_data:
#write tweets and its matrix to vectors.csv
with open('./vectors.csv', 'w') as vectors_data:
writer = csv.writer(vectors_data)
#write headr row to vectors.csv
writer.writerow([
"politician_name", "party", "tweet", "matrix",
"percentageOfMissingWords"
])
#rows to be written to vectors.csv
out_rows = []
reader = csv.DictReader(train_data)
for row in reader:
current_tweet = Tweet(row['tweet'])
#preprocess the tweet, and get list of tokens
class RelationVectorizer():
def __init__(self, word2vec_model_path, max_tokens_length, position_vector=True, word_position_size=10, ner=False, pos=False, dependency=False):
self.position_vector = position_vector
# word_position vectors will be filled when calling init_size function
self.word_position = None
self.word_position_size = word_position_size
self.sentence_vectorizer = WordVectorizer(word2vec_model_path, ner=ner, pos=pos, dependency=dependency)
# sizes of the output sequence matrix m is number of words in the sequence
# n is the size of the vector representation of each word in the sequence
self.m = max_tokens_length
self.n = self.sentence_vectorizer.model.vector_size + 2 * self.word_position_size
# original index = -l+1,....,0,...l-1
# array index = 0,.......,(l-1),...(2xl)-1
self.word_position = np.random.rand((2 * max_tokens_length) - 1, self.word_position_size)
def tokens_to_vec(self, tokens):
sentence_vec = []
for token in tokens:
vec = self.sentence_vectorizer.word2vec(token)
if len(vec.shape) == 0:
print 'length vec == 0'
return []
sentence_vec.append(vec)
sentence_vec = np.array(sentence_vec, dtype=np.float32)
return sentence_vec
def transform(self, sentence_matrix, labels):
sentence_matrix_out = np.zeros([0, self.m, self.n], np.float32)
valid_label = []
count = 0
for sentence_elements, label in tqdm(zip(sentence_matrix, labels)):
count += 1
sentence_vec = self.tokens_to_vec(sentence_elements["tokens"])
if sentence_vec == []:
print 'line %d sentence vector is null' % count
continue
entity1_vec = self.lookup_word_pos(sentence_elements["ent1_pos"], self.m) # dimension m x _
entity2_vec = self.lookup_word_pos(sentence_elements["ent2_pos"], self.m) # dimension m x _
pad_size = self.m - sentence_vec.shape[0]
if pad_size > 0:
temp = np.zeros((pad_size, self.sentence_vectorizer.model.vector_size))
sentence_vec = np.vstack([sentence_vec, temp])
# merging different parts of vector representation of words
sentence_matrix_vec = np.hstack([sentence_vec, entity1_vec, entity2_vec])
sentence_matrix_out = np.append(sentence_matrix_out, [sentence_matrix_vec], axis=0)
valid_label.append(label)
return sentence_matrix_out, valid_label
def lookup_word_pos(self, p, sentence_length):
"""
:param p: position of entity
:return: array of dimension self.m x self.word_position_size
example : if ent1 = 2 self.m = 10 i.e. : (w0, w1, w2(e1), w3, w4, w5, w6, w7, w8, w9)
return: word_position[-2:8] ===
add (l-1) to get indices between (0,2l-1) ===> word_position[7:17]
"""
start = -p + sentence_length - 1
end = start + sentence_length
return self.word_position[start:end]
