自然语言预处理

#英文预处理

#按空格进行分词，同时针对推文一些特性，去除@用户名，保留表情等一些特殊符号
tokenizer = TweetTokenizer()
for counter,rev in enumerate(reviews):
    # 去除HTML网页格式
    temp = BeautifulSoup(rev)
    text = temp.get_text()
    # 去除空格
    text = re.sub(' +',' ',text)
    test = re.sub(r'[()\[\]{}.,;:!?\<=>？@_^#$%"&*-],' ',text)
    # strip leading and trailing white space
    text = text.strip()
    # tokenize 
    tokens = tokenizer.tokenize(text)
    cleaned_reviews.append(tokens)
    if counter % round(len(reviews)/10) == 0:
        print(counter, '/', len(reviews), 'reviews cleaned')
# get list of tokens from all reviews
# 两个list变成一个list
all_tokens = [token for sublist in cleaned_reviews for token in sublist]
# 根据词频做index， 把单词转成index
counts = dict(Counter(all_tokens))
sorted_counts = sorted(counts.items(), key=operator.itemgetter(1), reverse=True)
# assign to each word an index based on its frequency in the corpus
# the most frequent word will get index equal to 1
word_to_index = dict([(tuple[0],idx+1) for idx, tuple in enumerate(sorted_counts)])
with open(path_to_IMDB + 'word_to_index_new.json', 'w') as my_file:
     json.dump(word_to_index, my_file, sort_keys=True, indent=4)

词共现矩阵的构建

https://github.com/urgedata/pythondata/blob/master/Text%20Analytics/ericbrown.ipynb 

#中文预处理

#jieba分词和去停用词
#jieba 分词可以将我们的自定义词典导入，格式 “词” “词性” “词频”
jieba.load_userdict('data/userdict.txt')
#定义一个keyword类
class keyword(object):
    def Chinese_Stopwords(self):          #导入停用词库
        stopword=[]
        cfp=open('data/stopWord.txt','r+','utf-8')   #停用词的txt文件
        for line in cfp:
            for word in line.split():
                stopword.append(word)
        cfp.close()
        return stopword
def Word_cut_list(self,word_str):
        #利用正则表达式去掉一些一些标点符号之类的符号。
        word_str = re.sub(r'\s+', ' ', word_str)  # trans 多空格 to空格
        word_str = re.sub(r'\n+', ' ', word_str)  # trans 换行 to空格
        word_str = re.sub(r'\t+', ' ', word_str)  # trans Tab to空格
        word_str = re.sub("[\s+\.\!\/_,$%^*(+\"\']+|[+——；！，”。《》，。：“？、~@#￥%……&*（）1234567①②③④)]+".decode("utf8"), "".decode("utf8"), word_str)
        wordlist = list(jieba.cut(word_str))#jieba分词
        wordlist_N = []
        chinese_stopwords=self.Chinese_Stopwords()
        for word in wordlist:
            if word not in chinese_stopwords:#词语的清洗：去停用词
                if word != '\r\n'  and word!=' ' and word != '\u3000'.decode('unicode_escape') \
                        and word!='\xa0'.decode('unicode_escape'):#词语的清洗：去全角空格
                    wordlist_N.append(word)
        return wordlist_N

#名词提取
def Word_pseg(self,word_str):  # 名词提取函数
        words = pseg.cut(word_str)
        word_list = []
        for wds in words:
            # 筛选自定义词典中的词，和各类名词，自定义词库的词在没设置词性的情况下默认为x词性，即词的flag词性为x
            if wds.flag == 'x' and wds.word != ' ' and wds.word != 'ns' \
                    or re.match(r'^n', wds.flag) != None \
                            and re.match(r'^nr', wds.flag) == None:
                word_list.append(wds.word)
        return word_list

import tensorflow.contrib.keras as kr

def read_file(filename):
    """读取文件数据"""
    contents, labels = [], []
    with open_file(filename) as f:
        for line in f:
            try:
                label, content = line.strip().split('\t')
                if content:
                    contents.append(list(content))#通过list把一句话分成一个个字
                    labels.append(native_content(label))
            except:
                pass
    return contents, labels


def build_vocab(train_dir, vocab_dir, vocab_size=5000):
    """根据训练集构建词汇表，存储"""
    data_train, _ = read_file(train_dir)
    all_data = []
    for content in data_train:
        all_data.extend(content)

    counter = Counter(all_data)
    count_pairs = counter.most_common(vocab_size - 1) #输出几个出现次数最多的元素
    words, _ = list(zip(*count_pairs)) #通过zip只取出其中的单词
    # 添加一个 <PAD> 来将所有文本pad为同一长度
    words = ['<PAD>'] + list(words)
    open_file(vocab_dir, mode='w').write('\n'.join(words) + '\n')

def read_vocab(vocab_dir):
    """读取词汇表"""
    # words = open_file(vocab_dir).read().strip().split('\n')
    with open_file(vocab_dir) as fp:
        # 如果是py2 则每个值都转化为unicode
        words = [native_content(_.strip()) for _ in fp.readlines()]
    word_to_id = dict(zip(words, range(len(words)))) 
    return words, word_to_id

def process_file(filename, word_to_id, cat_to_id, max_length=600):
    """将文件转换为id表示"""
    contents, labels = read_file(filename)
    data_id, label_id = [], []
    for i in range(len(contents)):
        data_id.append([word_to_id[x] for x in contents[i] if x in word_to_id])
        label_id.append(cat_to_id[labels[i]])
    # 使用keras提供的pad_sequences来将文本pad为固定长度
    x_pad = kr.preprocessing.sequence.pad_sequences(data_id, max_length)
    y_pad = kr.utils.to_categorical(label_id, num_classes=len(cat_to_id))  # 将标签转换为one-hot表示
    return x_pad, y_pad

#建立词表
text = open(path,encoding='utf-8').read().lower()
chars = set(text)
print ('total chars:', len(chars))
char_indices = dict((c, i) for i, c in enumerate(chars))
indices_char = dict((i, c) for i, c in enumerate(chars))
#kreas下运行LSTM的Input生成，在建立词表的基础上，数据向量化
print('Vectorization...')
X = np.zeros((len(sentences), maxlen, len(chars)), dtype=np.bool)
print(X)
y = np.zeros((len(sentences), len(chars)), dtype=np.bool)
for i, sentence in enumerate(sentences):
    for t, char in enumerate(sentence):
        X[i, t, char_indices[char]] = 1
    y[i, char_indices[next_chars[i]]] = 1

# 过滤词长，过滤停用词，只保留中文
def is_fine_word(word, min_length=2):
    rule = re.compile(r"^[\u4e00-\u9fa5]+$")
    if len(word) >= min_length and word not in STOP_WORDS and re.search(rule, word):
        return True
    else:
        return False

 

#逐字切分的处理方式，同时去掉一些常见的虚词，如“之”、“乎”、“者”、“也”。
def singCut(text):
     tex = [i.strip('\n').strip('\r').strip('。').strip('，|）|：|{|}|“|” |（|\n') for i in text]
     return list(filter(None, tex)) #去掉空字符
text = '云横秦岭家何在，雪拥蓝关马不前'

#虚词通用词库
stopwords = '而|何|乎|乃|且|其|若|所|为'
#去掉标点
poem = [[i.strip(') |: |?|{|}| “|” （| \n\n\r|。') for i in tex if i not in stopwords]for tex in text]
poem = list(filter(None, poem ))

　　

预处理（去特殊符号、去停用词、分词）　　

把词转成index（word to index）, 把原文都变成数值

去掉topN词频的以及小于TOPM词频的

对每篇进行 truncation and padding

word2vec训练 得到 w2v_model[word] 的embedding，加入CNN作为初始值（kreas里面训练需要把每个词转成embedding这种）

训练CNN模型

https://github.com/Tixierae/deep_learning_NLP

 

构建词汇表

categories转成id, 读取词汇表，构建word_to_id字典（字符级别）

读入训练数据，预处理，将文本pad到固定长度

批次训练CNN（tensorflow内部会自动初始化embedding）

预测

 https://github.com/gaussic/text-classification-cnn-rnn

 

 

 引用链接：

https://www.jianshu.com/p/aea87adee163