NLP实战——TextCNN实现新闻文本分类

本文介绍一个NLP中的常用模型——TextCNN。数据集来自天池新闻文本分类比赛数据，这里的文本已经经过匿名处理，可以到天池比赛地址下载数据集和查看模型在测试集上最终表现。

import tensorflow as tf
import pandas as pd
import matplotlib.pyplot as plt
from tensorflow.keras import layers
from sklearn.model_selection import train_test_split
from collections import Counter 
from gensim.models import Word2Vec
import numpy as np
from sklearn import preprocessing
from sklearn.metrics import precision_score, accuracy_score,recall_score, f1_score,roc_auc_score, precision_recall_fscore_support, roc_curve, classification_report

train_data=pd.read_csv("train_set.csv",sep='\t')
test_data=pd.read_csv("test_a.csv",sep='\t')

x=train_data['text']
y=train_data['label']
y=y.values.reshape(y.shape[0],1)

One-Hot编码

#y=pd.get_dummies(y)

enc = preprocessing.OneHotEncoder(categories='auto')
enc.fit(y)
y=enc.transform(y).toarray()
enc.categories_

[array([ 0,  1,  2,  3,  4,  5,  6,  7,  8,  9, 10, 11, 12, 13])]

分词

因为是匿名处理过的数据，所以这里的分词做的很简单，直接按空格进行拆分。

x=x.apply(lambda a:a.split())

训练word2vec词嵌入模型

w2v = Word2Vec(sentences=x,size=300,window=5,min_count=5,sample=1e-3,sg=1)
#window 句子中当前单词与预测词之间的最大距离。
#size 词向量的维数
#min_count 忽略总频率低于此的所有单词。
#sg 训练算法：1为skip-gram，否则为CBOW.
#hs 如果是1，则使用层次化的Softmax进行模型训练。如果为0或负数，则采用负采样.
#cbow_mean 如果为0，则使用上下文单词向量的和。如果是1，使用平均数，只有在使用CBOW时才适用。
#alpha 初始学习率。
#sample 配置高频词随机下采样的阈值，有效范围是(0, 1e-5)。

w2v.save("bag") #训练完成后保持模型到本地，一次训练多次使用；

# w2v = Word2Vec.load("bag") 第一次训练完成后，后续只需使用加载本地模型即可
def average(text,size=300):
    if len(text) < 1:
        return np.zeros(size)
    a = [w2v.wv[w] if w in w2v.wv else np.zeros(size) for w in text]
    length = len(a)
    summed = np.sum(a,axis=0)
    ave = np.divide(summed,length)
    return ave
x = x.apply(average)
list_corpus = x.tolist()

x_train,x_test,y_train,y_test = train_test_split(list_corpus,y,test_size=0.05,random_state=1)

x_train=pd.DataFrame(x_train)
x_test=pd.DataFrame(x_test)
x_train=x_train.values.reshape(x_train.shape[0],x_train.shape[1],1)
x_test=x_test.values.reshape(x_test.shape[0],x_test.shape[1],1)
y_train=np.array(y_train)
y_test=np.array(y_test)

x_train.shape,y_train.shape,x_test.shape,y_test.shape

((190000, 300, 1), (190000, 14), (10000, 300, 1), (10000, 14))

num_classes=14 # label中总共14个类

cnn=tf.keras.Sequential([tf.keras.layers.Conv1D(input_shape=x_train.shape[1:],filters=32,kernel_size=5,strides=1,
                                                padding='same',activation='relu',name='conv1'),
                           tf.keras.layers.MaxPool1D(pool_size=5,strides=2,name='pool1'),
                           tf.keras.layers.Conv1D(filters=64,kernel_size=5,strides=1,
                                                  padding='same',activation='relu',name='conv2'),
                           tf.keras.layers.MaxPool1D(pool_size=5,strides=2,name='pool2'),                           
                           tf.keras.layers.Conv1D(filters=128,kernel_size=5,strides=1,
                                                  padding='same',activation='relu',name='conv3'),
                           tf.keras.layers.MaxPool1D(pool_size=5,strides=2,name='pool3'),
                           tf.keras.layers.Flatten(),
                           tf.keras.layers.Dense(128,activation='relu',name='fc1'),
                           tf.keras.layers.Dense(64,activation='relu',name='fc2'),
                           tf.keras.layers.Dropout(0.5),
                           tf.keras.layers.Dense(32,activation='relu',name='fc3'),                        
                           tf.keras.layers.Dense(num_classes,activation='softmax',name='output')])
nadam = tf.keras.optimizers.Nadam(lr=1e-3)
cnn.compile(optimizer=nadam,loss='categorical_crossentropy',metrics=['categorical_accuracy']) #设置优化器和损失函数
cnn.summary()#查看模型基本信息

Model: "sequential"
_________________________________________________________________
Layer (type)                 Output Shape              Param #   
=================================================================
conv1 (Conv1D)               (None, 300, 32)           192       
_________________________________________________________________
pool1 (MaxPooling1D)         (None, 148, 32)           0         
_________________________________________________________________
conv2 (Conv1D)               (None, 148, 64)           10304     
_________________________________________________________________
pool2 (MaxPooling1D)         (None, 72, 64)            0         
_________________________________________________________________
conv3 (Conv1D)               (None, 72, 128)           41088     
_________________________________________________________________
pool3 (MaxPooling1D)         (None, 34, 128)           0         
_________________________________________________________________
flatten (Flatten)            (None, 4352)              0         
_________________________________________________________________
fc1 (Dense)                  (None, 128)               557184    
_________________________________________________________________
fc2 (Dense)                  (None, 64)                8256      
_________________________________________________________________
dropout (Dropout)            (None, 64)                0         
_________________________________________________________________
fc3 (Dense)                  (None, 32)                2080      
_________________________________________________________________
output (Dense)               (None, 14)                462       
=================================================================
Total params: 619,566
Trainable params: 619,566
Non-trainable params: 0
_________________________________________________________________

history=cnn.fit(x_train,y_train,epochs=100,verbose=1)

Train on 190000 samples
Epoch 1/100
WARNING:tensorflow:Entity .initialize_variables at 0x7ff50660a730> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: No module named 'tensorflow_core.estimator'
WARNING: Entity .initialize_variables at 0x7ff50660a730> could not be transformed and will be executed as-is. Please report this to the AutoGraph team. When filing the bug, set the verbosity to 10 (on Linux, `export AUTOGRAPH_VERBOSITY=10`) and attach the full output. Cause: No module named 'tensorflow_core.estimator'
190000/190000 [==============================] - 107s 563us/sample - loss: 0.7070 - categorical_accuracy: 0.7866
Epoch 2/100
190000/190000 [==============================] - 106s 557us/sample - loss: 0.4536 - categorical_accuracy: 0.8690
Epoch 3/100
190000/190000 [==============================] - 106s 560us/sample - loss: 0.4099 - categorical_accuracy: 0.8806
Epoch 4/100
190000/190000 [==============================] - 106s 559us/sample - loss: 0.3859 - categorical_accuracy: 0.8880
Epoch 5/100
190000/190000 [==============================] - 106s 559us/sample - loss: 0.3692 - categorical_accuracy: 0.8920
Epoch 6/100
190000/190000 [==============================] - 106s 559us/sample - loss: 0.3575 - categorical_accuracy: 0.8947
Epoch 7/100
190000/190000 [==============================] - 106s 559us/sample - loss: 0.3489 - categorical_accuracy: 0.8972
Epoch 8/100
190000/190000 [==============================] - 106s 558us/sample - loss: 0.3406 - categorical_accuracy: 0.8994
Epoch 9/100
190000/190000 [==============================] - 106s 559us/sample - loss: 0.3317 - categorical_accuracy: 0.9015
Epoch 10/100
190000/190000 [==============================] - 106s 
Epoch 20/100
190000/190000 [==============================] - 106s 559us/sample - loss: 0.2886 - categorical_accuracy: 0.9137
Epoch 30/100
190000/190000 [==============================] - 106s 558us/sample - loss: 0.2677 - categorical_accuracy: 0.9204
Epoch 40/100
190000/190000 [==============================] - 106s 558us/sample - loss: 0.2514 - categorical_accuracy: 0.9250
Epoch 50/100
190000/190000 [==============================] - 106s 558us/sample - loss: 0.2391 - categorical_accuracy: 0.9290
Epoch 60/100
190000/190000 [==============================] - 106s 559us/sample - loss: 0.2313 - categorical_accuracy: 0.9314
Epoch 70/100
190000/190000 [==============================] - 106s 558us/sample - loss: 0.2273 - categorical_accuracy: 0.9331
Epoch 80/100
190000/190000 [==============================] - 130s 684us/sample - loss: 0.2232 - categorical_accuracy: 0.9344
Epoch 90/100
190000/190000 [==============================] - 130s 685us/sample - loss: 0.2116 - categorical_accuracy: 0.9382
Epoch 100/100
190000/190000 [==============================] - 130s 686us/sample - loss: 0.2072 - categorical_accuracy: 0.9399

plt.plot(history.history['loss'])
plt.legend(['training'], loc='upper left')
plt.show()

测试集评估

# sortmax 结果转 onehot
def props_to_onehot(props):
    if isinstance(props, list):
        props = np.array(props)
    a = np.argmax(props, axis=1)
    b = np.zeros((len(a), props.shape[1]))
    b[np.arange(len(a)), a] = 1
    return b

# 自定义评估方法
def evalute_ty(predict_prop,predict_y,real_y):
    '''自定义模型评估方法
        评估不同阈值下模型的性能，将打印acc/precision/recall/f1/auc值
    Args:
        predict_prop:预测的概率值
        predict_y：预测值
        real_y ：真实标签
    '''
    acc=round(accuracy_score(real_y,predict_y),3)
    precision=round(precision_score(real_y, predict_y,average='weighted'),3)#准确率：预测为正的样本中实际为正的比例
    recall=round(recall_score(real_y, predict_y,average='weighted'),3)#召回率：所有正样本中预测为正的比例
    f1=round(f1_score(real_y, predict_y,average='weighted'),3)
    auc=round(roc_auc_score(y_test, predict_prop,multi_class='ovo'),3)
    print('acc {},precision {},recall {},f1 {},auc {}'.format(acc,precision,recall,f1,auc)) 

predict_prop=cnn.predict(x_test)
predict_prop.shape

(10000, 14)

predict_y=enc.inverse_transform(props_to_onehot(predict_prop))
real_y=enc.inverse_transform(y_test)

evalute_ty(predict_prop,predict_y,real_y)

acc 0.905,precision 0.906,recall 0.905,f1 0.905,auc 0.99

生成最终结果

x_test=test_data['text']
x_test=x_test.apply(lambda a:a.split())
x_test = x_test.apply(average) #注意此处的df["text"] 未分词
test_list_corpus = x_test.tolist()
x_test=pd.DataFrame(test_list_corpus)
x_test=x_test.values.reshape(x_test.shape[0],x_test.shape[1],1)
x_test.shape

(50000, 300, 1)

res=cnn.predict(x_test)
final=enc.inverse_transform(props_to_onehot(res))
final=final.reshape(final.shape[0],)

df=pd.DataFrame()
df['label']=final
df.to_csv('textcnn_submit.csv',index=None)

最终得分：