『TensotFlow』RNN中文文本_上

时间:2023-03-09 02:37:31
『TensotFlow』RNN中文文本_上

中文文字预处理流程

  • 文本处理
    • 读取+去除特殊符号
    • 按照字段长度排序
  • 辅助数据结构生成
    • 生成 {字符:出现次数} 字典
    • 生成按出现次数排序好的字符list
    • 生成 {字符:序号} 字典
    • 生成序号list

文本预处理生成字典是需要去重的,一般的思路是使用set数据结构来达成,不过这里使用的是collection.Counter,可以去重还能计数

这里的文本以全唐诗为例,一般一行为1首,目的是去掉作者,生成为“[诗主体]”的格式作为RNN输入,为了保证等长,引入字符“_”在后续处理中为长度不够的诗句补齐长度

之后生成对应的向量格式,首先做好{字符:序号} 字典并根据它生成“[序号诗]”作为向量化输入的准备

import numpy as np
import tensorflow as tf
from collections import Counter poetry_file = 'poetry.txt' poetrys = []
with open(poetry_file, 'r', encoding='utf-8') as f:
for line in f:
try:
title, content = line.strip().split(':')
content = content.replace(' ','') # 去空格,实际上没用到
if '_' in content or '(' in content or '(' in content or '《' in content or '[' in content:
continue
if len(content) < 5 or len(content) > 79:
continue
content = '[' + content + ']'
poetrys.append(content)
except Exception as e:
pass # 依照每个元素的长度排序
poetrys = sorted(poetrys, key=lambda poetry: len(poetry))
print('唐诗数量:', len(poetrys)) # 统计字出现次数
all_words = []
for poetry in poetrys:
all_words += [word for word in poetry]
counter = Counter(all_words)
print(counter.items())
# item会把字典中的每一项变成一个2元素元组,字典变成大list
count_pairs = sorted(counter.items(), key=lambda x:-x[1])
# 利用zip提取,因为是原生数据结构,在切片上远不如numpy的结构灵活
words, _ = zip(*count_pairs)
print(words) words = words[:len(words)] + (' ',) # 后面要用' '来补齐诗句长度
print(words)
# 转换为字典
word_num_map = dict(zip(words, range(len(words))))
# 把诗词转换为向量
to_num = lambda word: word_num_map.get(word, len(words))
poetry_vector = [list(map(to_num, poetry)) for poetry in poetrys]

生成RNN的batch数据,并生成标签,在这里使用了上面提到的'_'对诗句进行补齐(原因是RNN输入长度是固定的),

batch_size = 1
n_chunk = len(poetry_vector) // batch_size
x_batches = []
y_batches = []
for i in range(n_chunk):
start_index = i*batch_size
end_index = start_index + batch_size
batches = poetry_vector[start_index:end_index]
length = max(map(len, batches)) # 记录下最长的诗句的长度
xdata = np.full((batch_size, length), word_num_map[' '], np.int32)
for row in range(batch_size):
xdata[row,:len(batches[row])] = batches[row]
ydata = np.copy(xdata)
ydata[:,:-1] = xdata[:,1:]
"""
xdata ydata
[6,2,4,6,9] [2,4,6,9,9]
[1,4,2,8,5] [4,2,8,5,5]
"""
x_batches.append(xdata) # (n_chunk, batch, length)
y_batches.append(ydata)

由于本篇仅仅介绍预处理,所以下面列出向量化函数,这步处理之后的batch就可以作为RNN网络的输入了,

input_data = tf.placeholder(tf.int32, [batch_size, None])
output_targets = tf.placeholder(tf.int32, [batch_size, None]) embedding = tf.get_variable("embedding",[len(words),128])
inputs = tf.nn.embedding_lookup(embedding,input_data) sess = tf.Session()
sess.run(tf.global_variables_initializer())
print(sess.run(inputs,feed_dict={input_data: x_batches[0]}).shape)

解释一下 tf.nn.embedding_lookup,在之前cs231n的作业中做过类似的实现,就是把[batch [data]]映射为[batch [data [vactor]]],所以它需要提前生成一个映射用矩阵{总的字符数*RNN输入尺寸}。

『TensotFlow』RNN中文文本_下