def compute_vectors(self):
"""
Collect pas vector representations into a list.
:return: list of sentence vector representations.
"""
if not self.quiet:
print("Computing vectorial representation...")
vectors = np.zeros((1, 200, 134))
embeddings = sentence_embeddings(self.sentences)
centr_scores = centrality_scores(embeddings)
tf_idfs = tf_idf(self.sentences, os.getcwd() + "/data/idfs.dat")
# Position score, reference sentence length score, tf_idf, numerical data, centrality, title.
for j in range(min(len(self.sentences), 200)):
sent = self.sentences[j]
position_score = (len(self.sentences) - j) / len(self.sentences)
length_score = len(sent) / max(len(snt) for snt in self.sentences)
tf_idf_score = 0
numerical_score = 0
centrality_score = centr_scores[j]
title_sim_score = np.inner(np.array(embeddings[j]),
np.array(embeddings[-1]))
# Computing centrality and tf_idf score.
terms = list(set(stem_and_stopword(sent)))
for term in terms:
# Due to errors terms may be not present in the tf_idf dictionary.
if term in tf_idfs.keys():
tf_idf_score += tf_idfs[term]
else:
tf_idf_score += 0
if term.isdigit():
numerical_score += 1
# Some errors in the preprocessing may lead to zero terms, so it is necessary to avoid division by zero.
if len(terms):
tf_idf_score /= len(terms)
else:
tf_idf_score = 0
vectors[0, j, :] = np.append(
np.array([
position_score, length_score, tf_idf_score,
numerical_score, centrality_score, title_sim_score
]), np.array(embeddings[j]))
return vectors
def store_full_sentence_matrices(index, ref):
"""
Storing matrices for the extractive summarization task.
"""
if index < 0:
docs, references, _ = get_duc()
doc_path = "/dataset/duc/duc_doc_sent_matrix.dat"
ref_path = "/dataset/duc/duc_ref_sent_matrix.dat"
else:
docs_pas_lists, refs_pas_lists = get_pas_lists(index)
docs = get_sources_from_pas_lists(docs_pas_lists)
references = get_sources_from_pas_lists(refs_pas_lists)
dataset_path = "/dataset/nyt/" + str(index) + "/nyt" + str(index)
doc_path = dataset_path + "_doc_sent_matrix.dat"
ref_path = dataset_path + "_ref_sent_matrix.dat"
docs_no = len(docs) # First dimension, documents number.
# Second dimension, max document length (sparse), fixed in case of nyt.
max_sent_no = 200
# Third dimension, vector representation dimension.
sent_vec_len = 134
# The matrix are initialized as zeros, then they'll filled in with vectors for each docs' sentence.
refs_3d_matrix = np.zeros((docs_no, max_sent_no, sent_vec_len))
docs_3d_matrix = np.zeros((docs_no, max_sent_no, sent_vec_len))
# For each document the pas_list is extracted after cleaning the text and tokenizing it.
if ref:
doc_list = references
else:
doc_list = docs
for i in range(len(doc_list)):
doc = doc_list[i]
print("Processing doc " + str(i) + "/" + str(len(docs)))
doc = text_cleanup(doc)
# Splitting sentences (by dot).
sentences = tokens(doc)
embeddings = sentence_embeddings(sentences)
centr_scores = centrality_scores(embeddings)
tf_idfs = tf_idf(sentences, os.getcwd() + "/dataset/duc/duc_idfs.dat")
# Position score, reference sentence length score, tf_idf, numerical data, centrality, title.
for j in range(len(sentences)):
sent = sentences[j]
position_score = (len(sentences) - j) / len(sentences)
length_score = len(sent) / max(len(snt) for snt in sentences)
tf_idf_score = 0
numerical_score = 0
centrality_score = centr_scores[j]
title_sim_score = np.inner(np.array(embeddings[j]),
np.array(embeddings[-1]))
# Computing centrality and tf_idf score.
terms = list(set(stem_and_stopword(sent)))
for term in terms:
# Due to errors terms may be not present in the tf_idf dictionary.
if term in tf_idfs.keys():
tf_idf_score += tf_idfs[term]
else:
tf_idf_score += 0
if term.isdigit():
numerical_score += 1
# Some errors in the preprocessing may lead to zero terms, so it is necessary to avoid division by zero.
if len(terms):
tf_idf_score /= len(terms)
else:
tf_idf_score = 0
if ref:
refs_3d_matrix[i, j, :] = np.append([
position_score, length_score, tf_idf_score,
numerical_score, centrality_score, title_sim_score
], embeddings[j])
else:
docs_3d_matrix[i, j, :] = np.append([
position_score, length_score, tf_idf_score,
numerical_score, centrality_score, title_sim_score
], embeddings[j])
# Storing the matrices in the appropriate file, depending on the scoring system.
if ref:
with open(os.getcwd() + ref_path, "wb") as dest_f:
pickle.dump(refs_3d_matrix, dest_f)
else:
with open(os.getcwd() + doc_path, "wb") as dest_f:
pickle.dump(docs_3d_matrix, dest_f)
def get_important_words(corpus, text):
d = tf_idf(corpus, text)[0]['stats']
return sorted(d, key=d.get, reverse=True)[:max_words]
def main():
fileTotalManager = base.FileTotalManager('./file_lengths.json')
# load data into memory
print('loading csv data into memory...')
genome_tags = base.load_data(
'../data/genome-tags.csv', base.tags_adapter,
fileTotalManager.getFileTotal('genome-tags.csv'))
movies_info = base.load_data('../data/mlmovies.csv',
base.movie_info_adapter,
fileTotalManager.getFileTotal('mlmovies.csv'))
ratings_info = base.load_data(
'../data/mlratings.csv', base.RatingInfo,
fileTotalManager.getFileTotal('mlratings.csv'))
tags_info = base.load_data('../data/mltags.csv', base.TagInfo,
fileTotalManager.getFileTotal('mltags.csv'))
print('loading completed!')
# print(movie_actor[0].keys(), mltags[0].keys(), tags[0].keys(), mlmovies[0].keys(), mlusers[0].keys())
print('preprocessing data...')
# conversion
min_ts, max_ts = base.convert_timestamp(tags_info, 'timestamp')
# base.convert_timestamp(ratings_info, 'timestamp')
genome_tags = {k['tagId']: k['tag'] for k in genome_tags}
# movie_actor_list = base.get_moive_actor_list(movie_actor)
# genres_movie_list, min_yr, max_yr = base.get_genre_movies_list(movies_info)
movie_names = {k['movieid']: k['moviename'] for k in movies_info}
# actor_names = {k['id']: k['name'] for k in actor_info}
def tfidf_tag_weight(mr, ts):
return (1.0 / mr) * (ts - min_ts + 1) / (max_ts - min_ts + 1)
def no_weight(mr, ts):
return 1
print('building vectors')
# actor_tags_vector
# actors_tags_vector = base.actor_tag_vector(movie_actor, tags_info, no_weight)[1]
# actors_idf, actors_tfidf_tags_vector = base.actor_tag_vector(movie_actor, tags_info, tfidf_tag_weight)
# actors_idf = base.idf(actors_tfidf_tags_vector, actors_idf)
# for actor in actors_tfidf_tags_vector.keys():
# actors_tfidf_tags_vector[actor] = base.tf_idf(actors_tfidf_tags_vector[actor], actors_idf, 'tf-idf')
# movie_tags_vector
print('Building standard movie-tag vector')
movies_tags_vector = base.movie_tag_vector(movies_info, tags_info,
no_weight)[1]
print('\nBuilding tf-idf movie-tag vector')
movies_idf, movies_tfidf_tags_vector = base.movie_tag_vector(
movies_info, tags_info, tfidf_tag_weight)
movies_idf = base.idf(movies_tfidf_tags_vector, movies_idf)
for i, movie in enumerate(movies_tfidf_tags_vector.keys()):
movies_tfidf_tags_vector[movie] = base.tf_idf(
movies_tfidf_tags_vector[movie], movies_idf, 'tf-idf')
# movie_actors_vector
# movies_actors_vector = base.movie_actor_vector(movies_info, movie_actor, no_weight)[1]
# movies_actor_idf, movies_tfidf_actors_vector = base.movie_actor_vector(movies_info, movie_actor, tfidf_actor_weight)
# movies_actor_idf = base.idf(movies_tfidf_actors_vector, movies_actor_idf)
# for movie in movies_tfidf_actors_vector.keys():
# movies_tfidf_actors_vector[movie] = base.tf_idf(movies_tfidf_actors_vector[movie], movies_actor_idf, 'tf-idf')
# create actor-actor matrix
# actor_actor_similarity, actors_list, actors_index = build_actor_actor_matrix(actors_tfidf_tags_vector)
# create coactor-coactor matrix
# coactor_coactor_matrix, coactors_list, coactors_index = build_coactor_coactor_matrix(movie_actor)
# print('building AMY tensor')
# create Actor-Movie-Year tensor (AMY tensor)
# actor_movie_year_tensor, amy_tensor_info = build_actor_movie_year_tensor(movie_actor, movies_info)
print('\nbuilding TMR tensor')
# create Tag-Movie-Rating tensor (TMR tensor)
tag_movie_rating, tmr_tensor_info = build_tag_movie_rating_tensor(
genome_tags.keys(), ratings_info)
print('creating list')
# create watched list
users_watched_movies = base.get_users_watched_movies(
tags_info, ratings_info)
# create watched movies info
# watched_movies_info = base.get_moives_related_info(movies_info, ratings_info, movie_actor)
print('preprocessing completed!')
while True:
command_line = input('query>')
commands = command_line.split(' ')
relevance_feedback = None
if len(commands) > 0 and 'p3_task1' in commands[0]:
if len(commands) == 3:
if commands[2] == 'pf':
relevance_feedback = gen_prob_feedback_function(
movies_tags_vector)
else:
if not (commands[2] == 'PCA' or commands[2] == 'SVD'):
help()
continue
elif len(commands) == 4:
if commands[3] == 'pf':
relevance_feedback = gen_prob_feedback_function(
movies_tags_vector)
else:
help()
continue
WeightConstants.initialize(movie_names, tags_info, ratings_info)
if commands[0] == 'p3_task1a' and len(commands) > 2:
user_id = int(commands[1])
similarities = recommender_system_using_svd_pca(
user_id, users_watched_movies, movies_tfidf_tags_vector,
genome_tags, commands[2])
print_output_using(user_id, similarities, relevance_feedback)
elif commands[0] == 'p3_task1b' and len(commands) > 1:
user_id = int(commands[1])
similarities = recommender_system_using_lda(
user_id, users_watched_movies, movies_tags_vector, genome_tags)
print_output_using(user_id, similarities, relevance_feedback)
elif commands[0] == 'p3_task1c' and len(commands) > 1:
user_id = int(commands[1])
similarities = recommender_system_using_cp(
user_id, users_watched_movies, movies_tags_vector,
tag_movie_rating, tmr_tensor_info, genome_tags)
print_output_using(user_id, similarities, relevance_feedback)
elif commands[0] == 'p3_task1d' and len(commands) > 1:
user_id = int(commands[1])
similarities = recommender_system_using_ppr(
user_id, users_watched_movies, movies_tfidf_tags_vector)
print_output_using(user_id, similarities, relevance_feedback)
elif commands[0] == 'p3_task1e' and len(commands) > 1:
user_id = int(commands[1])
similarities = recommender_system_combining_all(
user_id, users_watched_movies, movies_tfidf_tags_vector,
movies_tags_vector, tag_movie_rating, tmr_tensor_info,
genome_tags)
print_output_using(user_id, similarities, relevance_feedback)
elif commands[0] == 'p3_task3' and len(commands) == 3:
lsh_indexing(genome_tags, movie_names, movies_tags_vector,
int(commands[1]), int(commands[2]))
elif commands[0] == 'p3_task5' and len(commands) > 1:
labelled_movies = {}
n = int(input("Enter number of labels: "))
while (n > 0):
label = input("Enter label: ")
movie_data = input("Enter space separated movies for label "
"" + label + ": ")
movies = movie_data.split(" ")
for i, m in enumerate(movies):
movies[i] = int(m)
labelled_movies[label] = movies
n -= 1
if commands[1] == 'NN' and len(commands) > 2:
recommender_system_for_labeling_movies(
movies_info, labelled_movies, genome_tags,
movies_tfidf_tags_vector, commands[1], int(commands[2]))
elif commands[1] == 'SVM' or commands[1] == 'DT':
recommender_system_for_labeling_movies(
movies_info, labelled_movies, genome_tags,
movies_tfidf_tags_vector, commands[1], 0)
elif len(commands) > 1 and (commands[0] == 'reset'
and commands[1] == 'wc'):
WeightConstants.reset()
print("WeightConstants data has been purged")
else:
help()
def extract_pas(sentences):
"""
Extracts the PASs from a list of sentences (
:param sentences: sentences from which to extract PAS.
"""
# Compute the TFIDF vector of all terms in the document.
tf_idfs = tf_idf(sentences, os.getcwd() + "/data/idfs.dat")
# Longest sentence length needed afterwards for the length score.
longest_sent_len = max(len(sent) for sent in sentences)
pas_list = []
for sent in sentences:
# Ignoring short sentences (errors).
if 3 < len(remove_punct(sent)) and len(sent) < 1000:
sent_index = sentences.index(sent)
# Substituting single apices with double apices to avoid errors with SRL.
sent = re.sub("\'([a-zA-Z0-9])([a-zA-Z0-9 ]+)([a-zA-Z0-9])\'",
r'" \1\2\3 "', sent)
annotations = _annotator.get_annoations(remove_punct(sent).split())
# Getting SRL annotations from SENNA.
sent_srl = annotations['srl']
# Getting POS tags from SENNA.
parts_of_speech = annotations['pos']
for raw_pas in sent_srl:
accept_pas = 1
out_of_order = 0
chk_sent = remove_punct(sent)
# Rejecting PASs with arguments that change the order (w.r.t. to the one of the original sentence);
# These represents the 10% of the total PASs and the 80% of them are incorrect.
for arg in raw_pas.values():
# Replacing double spaces with a single space to avoid some arguments to be ignored.
arg = remove_punct(arg.replace(" ", " "))
if chk_sent.find(arg) < 0:
accept_pas = False
out_of_order = 1
break
if accept_pas:
pas = Pas(sent, parts_of_speech, sent_index,
sent_srl.index(raw_pas), raw_pas, out_of_order)
pas_list.append(pas)
# Completing each PAS with its realization embeddings and vector representation.
# This process is done after the initialization as all the other pas are needed.
realized_pass = []
for pas in pas_list:
realized_pass.append(realize_pas(pas))
# Here the title is put together with the pass to avoid starting another embedding process
realized_pass.append(sentences[0])
pas_embeddings = sentence_embeddings(realized_pass)
# Get the centrality scores for the pas embeddings
pas_centralities = centrality_scores(pas_embeddings)
for pas in pas_list:
pas_index = pas_list.index(pas)
pas.complete_pas(
realized_pass[pas_index], pas_embeddings[pas_index],
len(sentences), longest_sent_len, tf_idfs,
pas_centralities[pas_index],
np.inner(np.array(pas_embeddings[pas_index]),
np.array(pas_embeddings[-1])))
return pas_list
#!python3
import text as text_parser
import utils
import xlsxwriter
import re
workbook = xlsxwriter.Workbook('tf-idf.xlsx')
print('Reading texts...')
all_texts = text_parser.get_text_corpus(9999, 'texts/news')
print('Done! Computing TF-IDF ranks...')
all_ranks = utils.tf_idf(all_texts)
print('\nDone! Writing results...')
text_no = 0
for dictionary in all_ranks:
text_no += 1
print('Writing worksheet ' + str(text_no) + '/' + str(len(all_ranks)) +
' (' + dictionary['title'][:10] + '...)',
end='\r')
sheet_name = str(text_no) + ' ' + re.sub('[\[\]:*?/\\\]', '',
dictionary['title'][:-1])
worksheet = workbook.add_worksheet(sheet_name[:28] +
('...' if len(sheet_name) > 28 else ''))
worksheet.write(0, 0, dictionary['title'])
worksheet.write(1, 0, '#')
worksheet.write(1, 1, 'Rank')
worksheet.write(1, 2, 'Word')
row = 2
sorted_dict = sorted(dictionary['stats'].items(),
key=lambda x: (x[1], x[0]),
reverse=True)
if __name__ == "__main__":
path = 'data/Data.csv'
tokenized_docs, ids = read_corpus(file_path=path,
has_tag=False,
has_id=True)
vocab_docs = list(Counter(chain(*tokenized_docs)).keys())
vocab, word2id, vectors = None, None, None
if not os.path.exists('word_vectors.npy'):
word2vec_model = KeyedVectors.load_word2vec_format(
'./GoogleNews-vectors-negative300.bin', binary=True)
save_word2vec_vocab(vocab_docs=vocab_docs,
word2vec_model=word2vec_model)
else:
vocab, word2id, vectors = load_word2vec_vocab()
tf_idf_matrix = tf_idf(tokenized_docs, vocab, word2id)
doc_vecs = get_document_vectors_word2vec(tokenized_docs, tf_idf_matrix,
word2id, vectors)
#model = kmeans(X_train=doc_vecs[:4000], n_clusters=4)
#doc_vecs = get_document_vectors_tfidf()
#model = kmeans(X_train=doc_vecs[:4000], n_clusters=4)
#predictions = model.predict(doc_vecs[4000:])
#model = gaussian_mixture_model(doc_vecs[:4000], n_components=4)
#print(model.predict(doc_vecs[4000:]))
print(hierarchical_clustering(X=doc_vecs[:4000], n_clusters=10))
rank_threshold = 0.09 # filter less relevant words, ranked in relative range 0..1
commons = set()
line = []
text = []
with open('tf_df_output.txt') as f:
content = f.readlines()
with open('files_count.txt') as f:
documents = int(f.read())
with open('file_name.txt') as f:
with open(f.read().strip()) as f2:
text = utils.get_normalized_words(f2.read())
tf_df_word = [x.strip().split() for x in content]
for [tf, df, word] in tf_df_word:
ranks[word] = utils.tf_idf(int(tf), int(df), documents)
if ranks[word] > max_rank:
max_rank = ranks[word]
for word in text:
rank = ranks[word] / max_rank
line.append(rank)
if rank < rank_threshold:
commons.add(word)
all_words = set(text) - commons
def add_to_hvg(w1, w2):
if w1 > w2:
temp = w1
def vectorize(self):
docs = [self.tokenizer(t.encode('utf-8')) for t in self.texts]
return utils.tf_idf(docs, normalize=True)
exit(1)
print('Reading all texts from /texts/news...')
corpus = get_text_corpus(root_dir=os.path.normpath('texts/news'))
main_text = None
for text in corpus:
if text['filename'] == file_path:
main_text = text
if not main_text:
print('A mystery duck found')
exit(1)
word_ranking = tf_idf(corpus, main_text)[0]
workbook = xlsxwriter.Workbook('article-rank.xlsx')
worksheet = workbook.add_worksheet(
re.sub('[\[\]:*?/\\\]', '', main_text['title'][0:28]))
worksheet.write(0, 0, main_text['title'])
worksheet.write(1, 0, '#')
worksheet.write(1, 2, 'Word')
worksheet.write(1, 1, 'Rank')
row = 2
for word in main_text['text']:
worksheet.write(row, 0, row - 1)
worksheet.write(row, 1, word)
worksheet.write(row, 2,
word_ranking['stats'][word] / word_ranking['max_rank'])
row += 1
workbook.close()
textds = TextDataSet(data_path='data/')
X_train, X_valid, X_test, y_train, y_valid, y_test = textds.generate_text_dataset(
train_size=500, valid_size=200, test_size=500)
print(X_train.shape, y_train.shape)
print(X_valid.shape, y_valid.shape)
print(X_test.shape, y_test.shape)
# print(textds.embedding_from_text(['I love coffee']).shape)
# %%
knn_ds = {
'hamming': ((X_train > 0).astype('float'), (X_valid > 0).astype('float')),
'euclidean': (X_train, X_valid),
'cosine': (tf_idf(X_train, alpha=1e-6,
beta=1e-9), tf_idf(X_valid, alpha=1e-6, beta=1e-9)),
}
best_acc = 0
best_metric = None
best_k = 0
for metric, (X_train_, X_valid_) in knn_ds.items():
for k in [1, 3, 5]:
clf = TextClassifier(Knn(n_neighbors=k, metric=metric))
clf.fit(X_train_, y_train)
acc = clf.score(X_valid_, y_valid)
print(metric, k, round(acc * 100, 2), sep=', ')
if acc > best_acc:
best_acc = acc
best_metric = metric
