def __init__(self):
self.tokenization = Tokenization(
import_module="jieba",
user_dict=config.USER_DEFINED_DICT_PATH,
chn_stop_words_dir=config.CHN_STOP_WORDS_PATH)
self.database = Database()
self.classifier = Classifier()
class TopicModelling(object):
def __init__(self):
self.tokenization = Tokenization(
import_module="jieba",
user_dict=config.USER_DEFINED_DICT_PATH,
chn_stop_words_dir=config.CHN_STOP_WORDS_PATH)
self.database = Database()
self.classifier = Classifier()
def create_dictionary(self,
raw_documents_list,
save_path=None,
is_saved=False):
"""
将文中每个词汇关联唯一的ID，因此需要定义词汇表
:param: raw_documents_list, 原始语料列表，每个元素即文本，如["洗尽铅华...", "风雨赶路人...", ...]
:param: savepath, corpora.Dictionary对象保存路径
"""
documents_token_list = []
for doc in raw_documents_list:
documents_token_list.append(self.tokenization.cut_words(doc))
_dict = corpora.Dictionary(documents_token_list)
# 找到只出现一次的token
once_items = [
_dict[tokenid] for tokenid, docfreq in _dict.dfs.items()
if docfreq == 1
]
# 在documents_token_list的每一条语料中，删除只出现一次的token
for _id, token_list in enumerate(documents_token_list):
documents_token_list[_id] = list(
filter(lambda token: token not in once_items, token_list))
# 极端情况，某一篇语料所有token只出现一次，这样该篇新闻语料的token列表就变为空，因此删除掉
documents_token_list = [
token_list for token_list in documents_token_list
if (len(token_list) != 0)
]
# 找到只出现一次的token对应的id
once_ids = [
tokenid for tokenid, docfreq in _dict.dfs.items() if docfreq == 1
]
# 删除仅出现一次的词
_dict.filter_tokens(once_ids)
# 消除id序列在删除词后产生的不连续的缺口
_dict.compactify()
if is_saved and save_path:
_dict.save(save_path)
logging.info(
"new generated dictionary saved in path -> {} ...".format(
save_path))
return _dict, documents_token_list
def renew_dictionary(self,
old_dict_path,
new_raw_documents_list,
new_dict_path=None,
is_saved=False):
documents_token_list = []
for doc in new_raw_documents_list:
documents_token_list.append(self.tokenization.cut_words(doc))
_dict = corpora.Dictionary.load(old_dict_path)
_dict.add_documents(documents_token_list)
if new_dict_path:
old_dict_path = new_dict_path
if is_saved:
_dict.save(old_dict_path)
logging.info(
"updated dictionary by another raw documents serialized in {} ... "
.format(old_dict_path))
return _dict, documents_token_list
def create_bag_of_word_representation(self,
raw_documents_list,
old_dict_path=None,
new_dict_path=None,
bow_vector_save_path=None,
is_saved_dict=False):
if old_dict_path:
# 如果存在旧的语料词典，就在原先词典的基础上更新，增加未见过的词
corpora_dictionary, documents_token_list = self.renew_dictionary(
old_dict_path, raw_documents_list, new_dict_path=new_dict_path)
else:
# 否则重新创建词典
start_time = time.time()
corpora_dictionary, documents_token_list = self.create_dictionary(
raw_documents_list,
save_path=new_dict_path,
is_saved=is_saved_dict)
end_time = time.time()
logging.info(
"there are {} mins spent to create a new dictionary ... ".
format((end_time - start_time) / 60))
# 根据新词典对文档(或语料)生成对应的词袋向量
start_time = time.time()
bow_vector = [
corpora_dictionary.doc2bow(doc_token)
for doc_token in documents_token_list
]
end_time = time.time()
logging.info(
"there are {} mins spent to calculate bow-vector ... ".format(
(end_time - start_time) / 60))
if bow_vector_save_path:
corpora.MmCorpus.serialize(bow_vector_save_path, bow_vector)
return documents_token_list, corpora_dictionary, bow_vector
@staticmethod
def transform_vectorized_corpus(corpora_dictionary,
bow_vector,
model_type="lda",
model_save_path=None):
# 如何没有保存任何模型，重新训练的情况下，可以选择该函数
model_vector = None
if model_type == "lsi":
# LSI(Latent Semantic Indexing)模型，将文本从词袋向量或者词频向量(更好)，转为一个低维度的latent空间
# 对于现实语料，目标维度在200-500被认为是"黄金标准"
model_tfidf = models.TfidfModel(bow_vector)
# model_tfidf.save("model_tfidf.tfidf")
tfidf_vector = model_tfidf[bow_vector]
model = models.LsiModel(tfidf_vector,
id2word=corpora_dictionary,
num_topics=config.TOPIC_NUMBER) # 初始化模型
model_vector = model[tfidf_vector]
if model_save_path:
model.save(model_save_path)
elif model_type == "lda":
model = models.LdaModel(bow_vector,
id2word=corpora_dictionary,
num_topics=config.TOPIC_NUMBER) # 初始化模型
model_vector = model[bow_vector]
if model_save_path:
model.save(model_save_path)
elif model_type == "tfidf":
model = models.TfidfModel(bow_vector) # 初始化
# model = models.TfidfModel.load("model_tfidf.tfidf")
model_vector = model[bow_vector] # 将整个语料进行转换
if model_save_path:
model.save(model_save_path)
return model_vector
def classify_stock_news(self,
unseen_raw_document,
database_name,
collection_name,
label_name="60DaysLabel",
topic_model_type="lda",
classifier_model="svm",
ori_dict_path=None,
bowvec_save_path=None,
is_saved_bow_vector=False):
historical_raw_documents_list = []
Y = []
for row in self.database.get_collection(database_name,
collection_name).find():
if label_name in row.keys():
if row[label_name] != "":
historical_raw_documents_list.append(row["Article"])
Y.append(row[label_name])
logging.info(
"fetch symbol '{}' historical news with label '{}' from [DB:'{}' - COL:'{}'] ... "
.format(collection_name, label_name, database_name,
collection_name))
le = preprocessing.LabelEncoder()
Y = le.fit_transform(Y)
logging.info(
"encode historical label list by sklearn preprocessing for training ... "
)
label_name_list = le.classes_ # ['中性' '利好' '利空'] -> [0, 1, 2]
# 根据历史新闻数据库创建词典，以及计算每个历史新闻的词袋向量；如果历史数据库创建的字典存在，则加载进内存
# 用未见过的新闻tokens去更新该词典
if not os.path.exists(ori_dict_path):
if not os.path.exists(bowvec_save_path):
_, _, historical_bow_vec = self.create_bag_of_word_representation(
historical_raw_documents_list,
new_dict_path=ori_dict_path,
bow_vector_save_path=bowvec_save_path,
is_saved_dict=True)
logging.info(
"create dictionary of historical news, and serialized in path -> {} ... "
.format(ori_dict_path))
logging.info(
"create bow-vector of historical news, and serialized in path -> {} ... "
.format(bowvec_save_path))
else:
_, _, _ = self.create_bag_of_word_representation(
historical_raw_documents_list,
new_dict_path=ori_dict_path,
is_saved_dict=True)
logging.info(
"create dictionary of historical news, and serialized in path -> {} ... "
.format(ori_dict_path))
else:
if not os.path.exists(bowvec_save_path):
_, _, historical_bow_vec = self.create_bag_of_word_representation(
historical_raw_documents_list,
new_dict_path=ori_dict_path,
bow_vector_save_path=bowvec_save_path,
is_saved_dict=True)
logging.info(
"historical news dictionary existed, which saved in path -> {}, but not the historical bow-vector"
" ... ".format(ori_dict_path))
else:
historical_bow_vec_mmcorpus = corpora.MmCorpus(
bowvec_save_path
) # type -> <gensim.corpora.mmcorpus.MmCorpus>
historical_bow_vec = []
for _bow in historical_bow_vec_mmcorpus:
historical_bow_vec.append(_bow)
logging.info(
"both historical news dictionary and bow-vector existed, load historical bow-vector to memory ... "
)
start_time = time.time()
updated_dictionary_with_old_and_unseen_news, unssen_documents_token_list = self.renew_dictionary(
ori_dict_path, [unseen_raw_document], is_saved=True)
end_time = time.time()
logging.info(
"renew dictionary with unseen news tokens, and serialized in path -> {}, "
"which took {} mins ... ".format(ori_dict_path,
(end_time - start_time) / 60))
unseen_bow_vector = [
updated_dictionary_with_old_and_unseen_news.doc2bow(doc_token)
for doc_token in unssen_documents_token_list
]
updated_bow_vector_with_old_and_unseen_news = []
updated_bow_vector_with_old_and_unseen_news.extend(historical_bow_vec)
updated_bow_vector_with_old_and_unseen_news.extend(unseen_bow_vector)
# 原先updated_bow_vector_with_old_and_unseen_news是list类型，
# 但是经过下面序列化后重新加载进来的类型是gensim.corpora.mmcorpus.MmCorpus
if is_saved_bow_vector and bowvec_save_path:
corpora.MmCorpus.serialize(
bowvec_save_path, updated_bow_vector_with_old_and_unseen_news
) # 保存更新后的bow向量，即包括新旧新闻的bow向量集
logging.info(
"combined bow vector(type -> 'list') generated by historical news with unseen bow "
"vector to create a new one ... ")
if topic_model_type == "lsi":
start_time = time.time()
updated_tfidf_model_vector = self.transform_vectorized_corpus(
updated_dictionary_with_old_and_unseen_news,
updated_bow_vector_with_old_and_unseen_news,
model_type="tfidf"
) # type -> <gensim.interfaces.TransformedCorpus object>
end_time = time.time()
logging.info(
"regenerated TF-IDF model vector by updated dictionary and updated bow-vector, "
"which took {} mins ... ".format((end_time - start_time) / 60))
start_time = time.time()
model = models.LsiModel(
updated_tfidf_model_vector,
id2word=updated_dictionary_with_old_and_unseen_news,
num_topics=config.TOPIC_NUMBER) # 初始化模型
model_vector = model[
updated_tfidf_model_vector] # type -> <gensim.interfaces.TransformedCorpus object>
end_time = time.time()
logging.info(
"regenerated LSI model vector space by updated TF-IDF model vector space, "
"which took {} mins ... ".format((end_time - start_time) / 60))
elif topic_model_type == "lda":
start_time = time.time()
model_vector = self.transform_vectorized_corpus(
updated_dictionary_with_old_and_unseen_news,
updated_bow_vector_with_old_and_unseen_news,
model_type="lda")
end_time = time.time()
logging.info(
"regenerated LDA model vector space by updated dictionary and bow-vector, "
"which took {} mins ... ".format((end_time - start_time) / 60))
# 将gensim.interfaces.TransformedCorpus类型的lsi模型向量转为numpy矩阵
start_time = time.time()
latest_matrix = corpus2dense(model_vector,
num_terms=model_vector.obj.num_terms).T
end_time = time.time()
logging.info(
"transform {} model vector space to numpy.adarray, "
"which took {} mins ... ".format(topic_model_type.upper(),
(end_time - start_time) / 60))
# 利用历史数据的话题模型向量(或特征)，进一步训练新闻分类器
start_time = time.time()
train_x, train_y, test_x, test_y = utils.generate_training_set(
latest_matrix[:-1, :], Y)
clf = self.classifier.train(train_x,
train_y,
test_x,
test_y,
model_type=classifier_model)
end_time = time.time()
logging.info(
"finished training by sklearn {} using latest {} model vector space, which took {} mins ... "
.format(classifier_model.upper(), topic_model_type.upper(),
(end_time - start_time) / 60))
label_id = clf.predict(latest_matrix[-1, :].reshape(1, -1))[0]
return label_name_list[label_id]
if __name__ == "__main__":
from Hisoka.classifier import Classifier
from Kite.database import Database
from sklearn import preprocessing
database = Database()
topicmodelling = TopicModelling()
raw_documents_list = []
Y = []
for row in database.get_collection("stocknews", "sz000001").find():
if "30DaysLabel" in row.keys():
raw_documents_list.append(row["Article"])
Y.append(row["30DaysLabel"])
le = preprocessing.LabelEncoder()
Y = le.fit_transform(Y)
_, corpora_dictionary, corpus = topicmodelling.create_bag_of_word_representation(
raw_documents_list)
model_vector = topicmodelling.transform_vectorized_corpus(
corpora_dictionary, corpus, model_type="lsi")
csr_matrix = utils.convert_to_csr_matrix(model_vector)
train_x, train_y, test_x, test_y = utils.generate_training_set(
csr_matrix, Y)
classifier = Classifier()
classifier.svm(train_x, train_y, test_x, test_y)
# lsi Tue, 15 Dec 2020 14:54:08 classifier.py[line:54] INFO train_pred: 0.9829 test_pred: 0.703 (只是去掉停用词、tab符以及空格符)
# lsi Tue, 15 Dec 2020 17:00:58 classifier.py[line:54] INFO train_pred: 0.9852 test_pred: 0.7492(去掉不含中文的词以及只有一个字符的词)
# lda Tue, 15 Dec 2020 17:29:56 classifier.py[line:54] INFO train_pred: 0.9498 test_pred: 0.7426(去掉不含中文的词以及只有一个字符的词)