In [1]:
import numpy as np
import sklearn
import matplotlib.pyplot as plt
import seaborn as sns;sns.set()
from sklearn.feature_extraction.text import TfidfVectorizer
import re
answers = []
questions = []
flag = 0
######################################
question = []
answer = []
file = open('xiaomi_raw.txt', 'r')
for line in file.readlines():
if line.startswith('用户问题'):
questions.append(line)
if answer:
answers.append(answer)
answer = []
##############
answer.append(line)
#############
elif len(line) < 6:
continue
else:
answer.append(line)
answers.append(answer)
file.close()
##########################################
######################################
question = []
answer = []
file1 = open('xiaomi_raw_2.txt', 'r')
for line in file1.readlines():
if line.startswith('用户问题'):
questions.append(line)
if answer:
answers.append(answer)
answer = []
##############
answer.append(line)
#############
elif len(line) < 6:
continue
else:
answer.append(line)
answers.append(answer)
file1.close()
##########################################
######################################
question = []
answer = []
file2 = open('xiaomi_raw_3.txt', 'r')
for line in file2.readlines():
if line.startswith('用户问题'):
questions.append(line)
if answer:
answers.append(answer)
answer = []
##############
answer.append(line)
#############
elif len(line) < 6:
continue
else:
answer.append(line)
answers.append(answer)
file2.close()
##########################################
answers = [''.join(ans) for ans in answers]
raw_questions = questions
raw_answers = answers
def get_accuracy(vectorizer, questions_vec, answers_vec):
"""
输入fit过的vectorizer，向量化的questions和answers，帮你返回正确率
"""
y = np.linspace(0, answers_vec.shape[0]-1, answers_vec.shape[0], dtype=np.int32)
y_predict = np.array([predict_answer(question_vec)[0] for question_vec in questions_vec])
return sum(y == y_predict) / answers_vec.shape[0], y, y_predict
def get_accuracy_with_threshold(vectorizer, questions_vec, answers_vec, threshold):
"""
输入一定的threshold，只有问题答案分值超过threshold才进行回答
Input:
threshold
Return:
correct_number
total_answer_number
accuracy
"""
correct = 0
total = 0
for i, question_vec in enumerate(questions_vec):
answer_idx, answer_scores = predict_answer(question_vec)
max_answer_score = np.max(answer_scores)
if max_answer_score < threshold:
# print(i, max_answer_score, 'max_answer_score < threshold', correct, total)
pass
else:
if answer_idx == i:
correct += 1
total += 1
print(i, np.max(answer_scores))
return correct, total, correct/total
def predict_answer(question_vec):
"""
Get answer choice from a single question_vec
Example:
predict_answer(vectorizer.transform(jieba.cut('你好什么是黑洞啊')))
Return:
最高分的回答，全部回答评分
"""
answer_scores = np.array([question_vec.multiply(answer_vec).sum() for answer_vec in answers_vec])
return np.argmax(answer_scores), answer_scores
def get_answer(question):
"""
提问题得到回答
Input:
中文问题
Return:
中文回答，全部回答评分
Example:
"""
question = filter_sentence(question)
question_vec = vectorizer.transform([question])
answer_score, answer_scores = predict_answer(question_vec)
# print(answer_score)
answer = answers[answer_score].replace(' ', '')
# print(1)
return answer, answer_scores
def check_accuracy_top_k(questions_vec, answers_vec, k, threshold=None):
"""
Input:
questions_vec:
Matrix representation of questions, with each question a row.
answers_vec:
Matrix representation of questions, with each answer a row.
k:
Top k answers are considered correct.
Return:
Accuracy of such metric.
"""
correct = 0
# total = questions_vec.shape[0]
total = 0
for i, question_vec in enumerate(questions_vec):
answer_scores = np.array([question_vec.multiply(answer_vec).sum() for answer_vec in answers_vec])
if threshold:
if max(answer_scores) < threshold:
continue
predict_ones = answer_scores.argsort()[-k:][::-1]
if i in predict_ones:
correct += 1
else:
print(questions[i])
for x in predict_ones:
print(answers[x], '<< wrong', answer_scores[x])
print(answers[i], '<< right', answer_scores[i])
total += 1
return correct/total
def filter_sentence(sentence):
clean = r'小米6'
new_sentence = re.sub(clean, '小米手机', sentence)
clean = r'小米电视4'
new_sentence = re.sub(clean, '小米电视', new_sentence)
clean = r'。|，|？|！|：|\n|:|\?'
new_sentence = re.sub(clean, '', new_sentence)
return new_sentence
In [2]:
questions = [filter_sentence(question) for question in questions]
answers = [filter_sentence(answer) for answer in answers]
import jieba
jieba.load_userdict('userdict.txt')
# questions = [' '.join(jieba.cut(question)) for question in questions]
# answers = [' '.join(jieba.cut(answer)) for answer in answers]
chinese_stopwords = '按,按照,俺,俺们,阿,别,别人,别处,别是,别的,别管,别说,不,不仅,不但,不光,不单,不只,不外乎,不如,不妨,不尽,不尽然,不得,不怕,不惟,不成,不拘,不料,不是,不比,不然,不特,不独,不管,不至于,不若,不论,不过,不问,比方,比如,比及,比,本身,本着,本地,本人,本,巴巴,巴,并,并且,非彼,彼时,彼此,便于,把,边,鄙人,罢了,被,般的,此间,此次,此时,此外,此处,此地,此,才,才能,朝,朝着,从,从此,从而,除非,除此之外,除开,除外,除了,除,诚然,诚如,出来,出于,曾,趁着,趁,处在,乘,冲,等等,等到,等,第,当着,当然,当地,当,多,多么,多少,对,对于,对待,对方,对比,得,得了,打,打从,的,的确,的话,但,但凡,但是,大家,大,地,待,都,到,叮咚,而言,而是,而已,而外,而后,而况,而且,而,尔尔,尔后,尔,二来,非独,非特,非徒,非但,否则,反过来说,反过来,反而,反之,分别,凡是,凡,个,个别,固然,故,故此,故而,果然,果真,各,各个,各位,各种,各自,关于具体地说,归齐,归,根据,管,赶,跟,过,该,给,光是,或者,或曰,或是,或则,或,何,何以,何况,何处,何时,还要,还有,还是,还,后者,很,换言之,换句话说,好,后,和,即,即令,即使,即便,即如,即或,即若,继而,继后,继之,既然,既是,既往,既,尽管如此,尽管,尽,就要,就算,就是说,就是了,就是,就,据,据此,接着,经,经过,结果,及,及其,及至,加以,加之,例如,介于,几时,几,截至,极了,简言之,竟而,紧接着,距,较之,较,进而,鉴于,基于,具体说来,兼之,借傥然,今,叫,将,可,可以,可是,可见,开始,开外,况且,靠,看,来说,来自,来着,来,两者,临,类如,论,赖以,连,连同,离,莫若,莫如,莫不然,假使,假如,假若,某,某个,某些,某某,漫说,没奈何,每当,每,慢说,冒,哪个,哪些,哪儿,哪天,哪年,哪怕,哪样,哪边,哪里,那里,那边,那般,那样,那时,那儿,那会儿,那些,那么样,那么些,那么,那个,那,乃,乃至,乃至于,宁肯,宁愿,宁可,宁,能,能否,你,你们,您,拿,难道说,内,哪,凭借,凭,旁人,譬如,譬喻,且,且不说,且说,其,其一,其中,其二,其他,其余,其它,其次,前后,前此,前者,起见,起,全部,全体,恰恰相反,岂但,却,去,若非,若果,若是,若夫,若,另,另一方面,另外,另悉,如若,如此,如果,如是,如同,如其,如何,如下,如上所述,如上,如,然则,然后,然而,任,任何,任凭,仍,仍旧,人家,人们,人,让,甚至于,甚至,甚而,甚或,甚么,甚且,什么,什么样,上,上下,虽说,虽然,虽则,虽,孰知,孰料,始而,所,所以,所在,所幸,所有,是,是以,是的,设使,设或,设若,谁,谁人,谁料,谁知,随着,随时,随后,随,顺着,顺,受到,使得,使,似的,尚且,庶几,庶乎,时候,省得,说来,首先,倘,倘使,倘或,倘然,倘若,同,同时,他,他人,他们们,她们,她,它们,它,替代,替,通过,腾,这里,这边,这般,这次,这样,这时,这就是说,这儿,这会儿,这些,这么点儿,这么样,这么些,这么,这个,这一来,这,正是,正巧,正如,正值,万一,为,为了,为什么,为何,为止,为此,为着,无论,无宁,无,我们,我,往,望,惟其,唯有,下,向着,向使,向,先不先,相对而言,许多,像,小,些,一,一些,一何,一切,一则,一方面,一旦,一来,一样,一般,一转眼,,由此可见,由此,由是,由于,由,用来,因而,因着,因此,因了,因为,因,要是,要么,要不然,要不是,要不,要,与,与其,与其说,与否,与此同时,以,以上,以为,以便,以免,以及,以故,以期,以来,以至,以至于,以致,己,已,已矣,有,有些,有关,有及,有时,有的,沿,沿着,于,于是,于是乎,云云,云尔,依照,依据,依,余外,也罢,也好,也,又及,又,抑或,犹自,犹且,用,越是,只当,只怕,只是,只有,只消,只要,只限,再,再其次,再则,再有,再者,再者说,再说,自身,自打,自己,自家,自后,自各儿,自从,自个儿,自,怎样,怎奈,怎么样,怎么办,怎么,怎,至若,至今,至于,至,纵然,纵使,纵令,纵,之,之一,之所以,之类,着呢,着,眨眼,总而言之,总的说来,总的来说,总的来看,总之,在于,在下,在,诸,诸位,诸如,咱们,咱,作为,只,最,照着,照,直到,综上所述,贼死,逐步,遵照,遵循,针对,致,者,则甚,则,咳,哇,哈,哈哈,哉,哎,哎呀,哎哟,哗,哟,哦,哩,矣哉,矣乎,矣,焉,毋宁,欤,嘿嘿,嘿,嘻,嘛,嘘,嘎登,嘎,嗳,嗯,嗬,嗡嗡,嗡,喽,喔唷,喏,喂,啷当,啪达,啦,啥,啐,啊,唉,哼唷,哼,咧,咦,咚,咋,呼哧,呸,呵呵,呵,呢,呜呼,呜,呗,呕,呃,呀,吱,吧哒,吧,吗,吓,兮,儿,亦,了,乎'
chinese_stopwords = chinese_stopwords.split(',')
# vocab = set((' '.join([answer for answer in answers] + [question for question in questions])).split())
vectorizer = TfidfVectorizer(ngram_range=(1,1), stop_words=chinese_stopwords, analyzer='word')
train_corpus = questions + answers
train_corpus = [i.replace('\n','') for i in train_corpus]
vectorizer.fit([' '.join(jieba.cut(sentence)) for sentence in train_corpus])
questions_vec = vectorizer.transform(questions)
answers_vec = vectorizer.transform(answers)
In [29]:
questions[0] = '用户问题小米手环的系统需求和硬件需求是什么 \
用户问题今天下雨了吗 \
用户问题小米电脑是什么配置啊 \
用户问题小米电视多少钱好不好用啊 \
用户问题小米扫地机器人怎么样 \
用户问题小米手机不发热了怎么办是不是坏了 \
用户问题小米电视弯曲了变成了曲面屏怎么办 \
用户问题小米雷军 are you ok \
用户问题小米手环会不会对人的身体有害啊 \
用户问题小米手环会不会影响今天的天气 \
用户问题小米电视会不会导致小孩沉迷看电视 \
用户问题小米手机下一代产品价格多少 \
用户问题小米手机会不会爆炸 \
用户问题小米电视发热大吗'
answers[0] = questions[0] + answers[0]
train_corpus = questions + answers
In [30]:
from sklearn.decomposition import LatentDirichletAllocation
lda_model = LatentDirichletAllocation(n_topics=25, max_iter=50,
learning_method='online',
learning_offset=50.,
random_state=0)
train_corpus_vec = vectorizer.fit_transform([' '.join(jieba.cut(sentence)) for sentence in train_corpus])
In [31]:
lda_model.fit(train_corpus_vec)
Out[31]:
In [32]:
def print_top_words(model, feature_names, n_top_words):
for topic_idx, topic in enumerate(model.components_):
print("Topic #%d:" % topic_idx)
print(" ".join([feature_names[i] for i in topic.argsort()[:-n_top_words - 1:-1]]))
print()
In [33]:
print_top_words(lda_model, vectorizer.get_feature_names(), 10)
In [34]:
lda_model.transform(vectorizer.transform([' '.join(jieba.cut(questions[0]))]))
Out[34]:
In [35]:
questions[0]
Out[35]:
In [36]:
lda_model.get_params
Out[36]:
In [41]:
answers[np.argmax(lda_model.transform(vectorizer.transform([('电视弯曲')])))]
Out[41]:
In [38]:
answers[0] = questions[0] + answers[0]
In [ ]: