Tensorflow之知识图谱---（1）NLP入门 评论情感分类

 本笔记结合书《TensorFlow知识图谱实战》--王晓华著

一、详细代码

import numpy as np
import tensorflow as tf

labels=[]
context=[]
vocab=set()

with open("ChnSentiCorp.txt",mode='r',encoding='utf-8') as emotion_file:
# with open("test.txt", mode='r', encoding='utf-8') as emotion_file:
    for line in emotion_file.readlines():
        line=line.strip().split(",")
        labels.append(int(line[0]))

        # print(line)
        # print(len(line))

        text=line[1]
        # print(text)
        # print(type(text))

        context.append(line[1])
        # print("context:",context)

        for char in text:
            vocab.add(char)
            # print(char)
            # print(vocab)
            # print(type(vocab))
vocab_list = list(vocab)
# print(vocab_list)
# print(len(vocab_list))
vocab_list = list(sorted(vocab))
# print(vocab_list)
# print(len(vocab_list))

token_list=[ ]
for text in context:
    token = [vocab_list.index(char) for char in text ]

    # 分解版本：
    # print("context:",context)
    # print("text:",text)
    # token=[ ]
    # for char in text:
    #     # print("char:",char)
    #     token.append(vocab_list.index(char))
        # print("token:",token)
    #以80个字符为长度对句子进行截取或补全
    # print("len(token);",len(token))
    # print("token[:80]+[0]*(80-len(token)):",token[:80]+[0]*(80-len(token)))
    token= token[:80]+[0]*(80-len(token))
    token_list.append(token)
# print("token_list:",token_list,"\n")

#格式化处理
token_list=np.array(token_list)
# print("token_list:",token_list,"\n")
# print(labels)
labels=np.array(labels)
# print(labels)


input_token = tf.keras.Input(shape=(80,))
embedding = tf.keras.layers.Embedding(input_dim=3508,output_dim=128)(input_token)
embedding = tf.keras.layers.Bidirectional(tf.keras.layers.GRU(128))(embedding)
output = tf.keras.layers.Dense(2,activation=tf.nn.softmax)(embedding)
model = tf.keras.Model(input_token,output)

model.compile(optimizer='adam',loss=tf.keras.losses.sparse_categorical_crossentropy,metrics=['accuracy'])

model.fit(token_list,labels,epochs=10,verbose=2)

 

二、关键思路代码解析

 1.读取文件 把句子用逗号分割为词组

for line in emotion_file.readlines():

        line=line.strip().split(",")

此时line以列表 [ ]存储

 

2. 遍历词组中的每个字符 并删除掉重复的set()

vocab=set()

for char in text:
        vocab.add(char)

• text是词组
• 遍历每个字符 并查看是否已经在vocab存在 不存在则append
• 此时的vocab是set()类型

3. 用文本的索引号数字来代替符号

简化代码：

for text in context:

        token= [ vocab_list.index(char) for char in text ]

思路详细解析：

for text in context:

token=[ ]  ---要先定义好为list  否则后面不能append
    for char in text:
        token.append(vocab_list.index(char))  ---前面排序好的

部分运行结果：

vocab_list：['!', '"', '2', 'M', '。', '一', '下', '不', '个', '也', '了', '人', '体', '公', '加', '务', '去', '品', '商', '多', '大', '太', '好', '实', '宽', '少', '川', '差', '床', '应', '店', '很', '惠', '感', '房', '整', '无', '早', '有', '本', '沙', '济', '的', '离', '经', '视', '觉', '论', '该', '距', '路', '身', '较', '边', '近', '这', '那', '酒', '重', '错', '问', '间', '题', '食', '餐', '，']

context: ['"距离川沙公路较近', '商务大床房，房间很大，床有2M宽，整体感觉经济实惠不错!', '早餐太差，无论去多少人，那边也不加食品的。酒店应该重视一下这个问题了。房间本身很好。']

text: "距离川沙公路较近

char: "

token: [1]

char: 距

token: [1, 49]

char: 离

token: [1, 49, 43]

4.以80个字符为长度对句子进行截取或补全

token= token[:80]+[0]*(80-len(token))

•  [0]*(80-len(token)：不足80个的剩余空用0补齐

• 例子：token长42 则需要补齐28个0

  token: [37, 64, 21, 27, 65, 36, 47, 16, 19, 25, 11, 65, 56, 53, 9, 7, 14, 63, 17, 42, 4, 57, 30, 29, 48, 58, 45, 5, 6, 55, 8, 60, 62, 10, 4, 34, 61, 39, 51, 31, 22, 4]

  token[:80]+[0]*(80-len(token)): [37, 64, 21, 27, 65, 36, 47, 16, 19, 25, 11, 65, 56, 53, 9, 7, 14, 63, 17, 42, 4, 57, 30, 29, 48, 58, 45, 5, 6, 55, 8, 60, 62, 10, 4, 34, 61, 39, 51, 31, 22, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

5. 把每个text（每短句）存入token_list --- 一句话一个token

token_list: （举例三句话）

[[1, 49, 43, 26, 40, 13, 50, 52, 54, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],

[18, 15, 20, 28, 34, 65, 34, 61, 31, 20, 65, 28, 38, 2, 3, 24, 65, 35, 12, 33, 46, 44, 41, 23, 32, 7, 59, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],

[37, 64, 21, 27, 65, 36, 47, 16, 19, 25, 11, 65, 56, 53, 9, 7, 14, 63, 17, 42, 4, 57, 30, 29, 48, 58, 45, 5, 6, 55, 8, 60, 62, 10, 4, 34, 61, 39, 51, 31, 22, 4, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]

 6.对数据集进行格式化处理 np.array( )

        np.array ( ) : 创建一个数组

token_list=np.array(token_list)

labels=np.array(labels)

结果：前：[1, 1, 1]   后：[1 1 1]

 

 三、代码运行结果

Epoch 1/10
243/243 - 10s - loss: 0.4815 - accuracy: 0.7691
Epoch 2/10
243/243 - 10s - loss: 0.3476 - accuracy: 0.8514
Epoch 3/10
243/243 - 11s - loss: 0.2802 - accuracy: 0.8840
Epoch 4/10
243/243 - 12s - loss: 0.2436 - accuracy: 0.8990
Epoch 5/10
243/243 - 11s - loss: 0.2051 - accuracy: 0.9138
Epoch 6/10
243/243 - 11s - loss: 0.1654 - accuracy: 0.9339
Epoch 7/10
243/243 - 11s - loss: 0.1399 - accuracy: 0.9449
Epoch 8/10
243/243 - 11s - loss: 0.1187 - accuracy: 0.9517
Epoch 9/10
243/243 - 11s - loss: 0.0924 - accuracy: 0.9643
Epoch 10/10
243/243 - 11s - loss: 0.0798 - accuracy: 0.9679