Relation Classification via Convolutional Deep Neural Network 模型实现

  Keras是最容易使用的深度学习框架了，在此使用该框架进行模型实现。其中Python版本为3x。

1. 导库

import numpy as np
import gzip
import sys
from keras.models import Model
from keras.layers import Dense, Input
from keras.layers import Dropout, GaussianNoise
from keras.layers import Embedding, concatenate
from keras.layers import Convolution1D, GlobalMaxPooling1D #, MaxPooling1D
from sklearn.model_selection import StratifiedKFold
from sklearn.metrics import classification_report
import pickle as pkl

2. 参数设置

# 设置初始化种子，从而保证结果的复现性。
seed = 666
np.random.seed(seed) # # fix random seed for reproducibility
# 模型超参数
filter_sizes = 3
num_filters = 100
max_num_words = 20000 #词典中词的个数
position_dims = 50

# 训练参数
batch_size = 64 
num_epochs = 1
dropout_rate = 0 #0.5
std_noise = 0 #0.3

3. 读取数据

print("Load dataset")
f = gzip.open('data.pkl.gz', 'rb')
data = pkl.load(f)
f.close()

sentence_train, yLabel_train, positionMatrix1_train, positionMatrix2_train = data['train']
sentence_test, yLabel_test, positionMatrix1_test, positionMatrix2_test = data['test']
max_position = max(np.max(positionMatrix1_train), np.max(positionMatrix2_train))+1
n_out = yLabel_train.shape[1]
max_sequence_length = sentence_train.shape[1]

print("Loading embeddings")
embedding_matrix = np.load(open('embeddings.npz', 'rb'))

Y = np.argmax(yLabel_train,axis=-1)
print("Training the model")

4. 创建和定义模型

words_input = Input(shape=(max_sequence_length,), dtype='int32', name='words_input')
words = Embedding(embedding_matrix.shape[0], embedding_matrix.shape[1], 
                  weights=[embedding_matrix], trainable=False)(words_input)

distance1_input = Input(shape=(max_sequence_length,), dtype='int32', name='distance1_input')
distance1 = Embedding(max_position, position_dims)(distance1_input)

distance2_input = Input(shape=(max_sequence_length,), dtype='int32', name='distance2_input')
distance2 = Embedding(max_position, position_dims)(distance2_input)

output = concatenate([words, distance1, distance2])

output = GaussianNoise(std_noise)(output)

output = Convolution1D(filters=num_filters,
                        kernel_size=filter_sizes,
                        padding='valid',
                        activation='relu',
                        strides=1)(output)
#    output = MaxPooling1D(5)(output)
#    output = Convolution1D(num_filters, filter_sizes, activation='relu')(output)
output = GlobalMaxPooling1D()(output)
output = Dropout(dropout_rate)(output)
output = Dense(n_out, activation='softmax')(output)
model = Model(inputs=[words_input, distance1_input, distance2_input], outputs=[output])
# compile the model
model.compile(optimizer='adam', loss='categorical_crossentropy', metrics=['accuracy'])

5. 交叉验证

# define 10-fold cross validation test harness
k = 10
skf = StratifiedKFold(n_splits=k, shuffle=True, random_state=seed)

j=1
cvscores = []
for train, test in skf.split(sentence_train,Y):
    # Fit the model
    print("train k:",j,"/",k)      
    model.fit([sentence_train[train], positionMatrix1_train[train],positionMatrix2_train[train]],
              yLabel_train[train], epochs=num_epochs, batch_size=batch_size, verbose=1)
    # evaluate the model
    scores = model.evaluate([sentence_train[test], positionMatrix1_train[test],positionMatrix2_train[test]],
                            yLabel_train[test], verbose=0)
    print("val_acc: %.2f%%" % (scores[1]*100)) #model.metrics_names[1]
    cvscores.append(scores[1] * 100)
    j = j+1

print("%.2f%% (+/- %.2f%%)" % (np.mean(cvscores), np.std(cvscores)))

predicted_yLabel=model.predict([sentence_test, positionMatrix1_test, positionMatrix2_test], 
                     batch_size=None, verbose=0, steps=None)

predicted_yLabel = np.argmax(predicted_yLabel,axis=-1)
yLabel_test = np.argmax(yLabel_test,axis=-1)