WSDm问题语义匹配分类算法，SentenceEmbedding实现，自然语言处理

优化版后的网络架构图&＃xff1a;

参考文献&＃xff1a;

https://arxiv.org/pdf/1808.08762.pdf

实现代码&＃xff08;自己修改了一部分&＃xff0c;性能更高&＃xff09;&＃xff1a;

# -*- coding: utf-8 -*-
"""
Created on Thu Nov 22 12:02:08 2018&＃64;author: Lenovo
"""import pandas as pd
from keras.layers import Subtract,Dense,Reshape,BatchNormalization,Lambda,Flatten,Dot,MaxPooling2D,AveragePooling2D,AveragePooling1D,Concatenate,MaxPooling1D,Conv2D,Conv1D,Embedding,CuDNNLSTM,Input,Activation,Multiply,Bidirectional,Dropout
from keras.models import Model,Sequential
from keras.optimizers import SGD,Adam
from keras.callbacks import ModelCheckpoint,EarlyStopping
from keras.preprocessing.text import Tokenizer
from keras.preprocessing.sequence import pad_sequences
from keras.utils.np_utils import to_categorical
import numpy as np
from data_process import get_data
from keras.utils import plot_model
import matplotlib.pyplot as plt
import keras.backend as K
import tensorflow as tfdef adbsub(x):x1,x2 &＃61; xreturn tf.abs(tf.subtract(x1,x2))def calc_01(x):x_1,x_2 &＃61; xreturn K.cast(K.equal(x_1,x_2),dtype&＃61;&＃39;float32&＃39;)def stack_dot_01(x):x_1,x_2 &＃61; xreturn K.stack(x,axis&＃61;-1)category &＃61; ["unrelated","agreed","disagreed"]
#train_x_1,train_x_2,label,vsize,test_x_1,test_x_2,test,train_jiao,test_jiao &＃61; get_data()x_1_input &＃61; Input(shape&＃61;([50]))
x_2_input &＃61; Input(shape&＃61;([50]))
x_jiao_in &＃61; Input(shape&＃61;([1]))embedder &＃61; Embedding(input_dim&＃61;vsize&＃43;1, output_dim&＃61;300)
x_1_in &＃61; embedder(x_1_input)
x_2_in &＃61; embedder(x_2_input)
#x_1 &＃61; Conv1D(32,3,strides&＃61;1,padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x_1)
#x_2 &＃61; Conv1D(32,3,strides&＃61;1,padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x_2)
##x_1 &＃61; BatchNormalization()(x_1)
##x_2 &＃61; BatchNormalization()(x_2)
###做这个池化 有可能会导致信息失真
#x_1 &＃61; MaxPooling1D(pool_size&＃61;3,strides&＃61;2)(x_1)
#x_2 &＃61; MaxPooling1D(pool_size&＃61;3,strides&＃61;2)(x_2)
#
bilstm_1 &＃61; Bidirectional(CuDNNLSTM(units&＃61;150,return_sequences&＃61;True,return_state&＃61;True))
lstm11 &＃61; bilstm_1(x_1_in)
x_1_b_m_1 &＃61; MaxPooling1D()(lstm11[0])
lstm12 &＃61; bilstm_1(x_2_in)
x_2_b_m_1 &＃61; MaxPooling1D()(lstm12[0])
#
bilstm_2 &＃61; Bidirectional(CuDNNLSTM(units&＃61;150,return_sequences&＃61;True,return_state&＃61;True))
lstm21 &＃61; bilstm_2(Multiply()([x_1_in,lstm11[0]]))
x_1_b_m_2 &＃61; MaxPooling1D()(lstm11[0])
lstm22 &＃61; bilstm_2(Multiply()([x_2_in,lstm12[0]]))
x_2_b_m_2 &＃61; MaxPooling1D()(lstm12[0])#
bilstm_3 &＃61; Bidirectional(CuDNNLSTM(units&＃61;150,return_sequences&＃61;True,return_state&＃61;True))
lstm31 &＃61; bilstm_3(Multiply()([x_1_in,lstm21[0]]))
x_1_b_m_3 &＃61; MaxPooling1D()(lstm31[0])
lstm32 &＃61; bilstm_3(Multiply()([x_2_in,lstm22[0]]))
x_2_b_m_3 &＃61; MaxPooling1D()(lstm32[0])x_Concatenate &＃61; Concatenate()([x_1_b_m_1,x_2_b_m_1])
x_Subtract &＃61; Lambda(adbsub)([x_1_b_m_2,x_2_b_m_2])
x_Multiply &＃61; Multiply()([x_1_b_m_3,x_2_b_m_3])
x_m_1 &＃61; Concatenate()([x_Concatenate,x_Subtract,x_Multiply])#x_m_2 &＃61; Concatenate()([x_2_b_m_1,x_2_b_m_2,x_2_b_m_3])#x_1 &＃61; BatchNormalization()(x_1)
#x_2 &＃61; BatchNormalization()(x_2)
##
#bilistm_merge &＃61; Multiply()([x_1,x_2])
#b_out &＃61; Bidirectional(CuDNNLSTM(units&＃61;128))(bilistm_merge)
#b_jiao &＃61; Bidirectional(CuDNNLSTM(units&＃61;128))(bilistm_merge)
#b_out &＃61; BatchNormalization()(b_out)
#
#bilstm_2 &＃61; Bidirectional(CuDNNLSTM(units&＃61;50))
#x_1 &＃61; bilstm_2(x_1)
#x_2 &＃61; bilstm_2(x_2)
#x_1 &＃61; BatchNormalization()(x_1)
#x_2 &＃61; BatchNormalization()(x_2)
#
#x_2c &＃61; Concatenate(axis&＃61;-1)([x_1,x_2])
#x_2c &＃61; BatchNormalization()(x_2c)
##
#x_2c &＃61; Reshape((2,100,1))(x_2c)
#x_2c &＃61; Conv2D(2,kernel_size&＃61;(2,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x_2c)
#x_2c &＃61; BatchNormalization()(x_2c)
#x_2c &＃61; Conv2D(2,kernel_size&＃61;(2,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x_2c)
#x_2c &＃61; BatchNormalization()(x_2c)
#
#x_2c &＃61; AveragePooling2D((2,2),(1,1))(x_2c)
#x_2c &＃61; Flatten()(x_2c)
#x_2c &＃61; Dense(256,activation&＃61;&＃39;relu&＃39;)(x_2c)
#print(x_2c.shape)
#x_1 &＃61; Reshape((1,100))(x_1)
#x_2 &＃61; Reshape((100,1))(x_2)
#x_01 &＃61; Lambda(calc_01)([x_1,x_2])
#x_dot &＃61; Multiply()([x_1,x_2])
#
#x_dot &＃61; Reshape((200,200,1))(x_dot)
#x_01 &＃61; Reshape((200,200,1))(x_01)
#
#x &＃61; Lambda(stack_dot_01)([x_dot,x_01])
#print(x.shape)
#x &＃61; Conv2D(16,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
#x &＃61; Conv2D(16,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
##x &＃61; BatchNormalization()(x)
#x &＃61; MaxPooling2D((2,2),(2,2))(x)
##
#x &＃61; Conv2D(32,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
#x &＃61; Conv2D(32,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
##x &＃61; BatchNormalization()(x)
#x &＃61; MaxPooling2D((2,2),(2,2))(x)
##
#x &＃61; Conv2D(32,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
#x &＃61; Conv2D(32,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
##x &＃61; BatchNormalization()(x)
##x &＃61; AveragePooling2D((2,2),(2,2))(x)
##x &＃61; Conv2D(64,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
##x &＃61; Conv2D(64,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
##x &＃61; BatchNormalization()(x)
#x &＃61; MaxPooling2D((2,2),(2,2))(x)
##
#
x &＃61; Flatten()(x_m_1)
x &＃61; Dropout(0.5)(x)
##x &＃61; BatchNormalization()(x)
##x &＃61; Multiply()([x_1,x_2])
##x &＃61; Activation(&＃39;relu&＃39;)(x)
x &＃61; Dense(256,activation&＃61;&＃39;relu&＃39;)(x)
#x_jiao_in_1 &＃61; Dense(5)(x_jiao_in)
#x_jiao_in_1 &＃61; Dense(1)(x_jiao_in)
x &＃61; Concatenate(axis&＃61;1)([x,x_jiao_in])#x &＃61; BatchNormalization()(x)
#x &＃61; Concatenate(axis&＃61;-1)([x,b_out])
#x &＃61; Conv2D(32,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
#x &＃61; Conv2D(32,kernel_size&＃61;(3,3),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
#x &＃61; MaxPooling2D((2,2),(2,2))(x)
#x &＃61; Reshape((3,256,1))(x)
#x &＃61; Conv2D(3,kernel_size&＃61;(3,256),strides&＃61;(1,1),padding&＃61;&＃39;valid&＃39;,activation&＃61;&＃39;relu&＃39;)(x)
##x &＃61; BatchNormalization()(x)
#x &＃61; Conv2D(3,kernel_size&＃61;(3,256),strides&＃61;(1,1),padding&＃61;&＃39;same&＃39;,activation&＃61;&＃39;softmax&＃39;)(x)
##x &＃61; BatchNormalization()(x)
#out &＃61; Reshape(([3]))(x)
#x &＃61; Flatten()(x)#x &＃61; Dropout(0.2)(x)print(&＃39;朴实无华网络__很猛&＃39;)
#x &＃61; Concatenate(axis&＃61;-1)([x_1,x_2])
#x &＃61; Flatten()(x)x &＃61; Dropout(0.2)(x)out &＃61; Dense(3,activation&＃61;&＃39;softmax&＃39;)(x)model &＃61; Model([x_1_input,x_2_input,x_jiao_in],[out])model.compile(optimizer&＃61;&＃39;adam&＃39;,loss&＃61;&＃39;categorical_crossentropy&＃39;,metrics&＃61;[&＃39;acc&＃39;])
model.summary()
plot_model(model,&＃39;model.png&＃39;)
mcp &＃61; ModelCheckpoint(&＃39;wsdm_duoronghe_jiyuci_best.h5&＃39;,save_best_only&＃61;True,monitor&＃61;&＃39;val_acc&＃39;,verbose&＃61;1)
es &＃61; EarlyStopping(patience&＃61;5,monitor&＃61;&＃39;loss&＃39;,verbose&＃61;1)
cl &＃61; [mcp,es]
history &＃61; model.fit(x&＃61;[train_x_1,train_x_2,train_jiao],y&＃61;label,class_weight&＃61;{0:1,1:5,2:10},batch_size&＃61;256,epochs&＃61;100,verbose&＃61;1,callbacks&＃61;cl,validation_split&＃61;0.1,shuffle&＃61;True)x&＃61;history.epoch
y&＃61;history.history[&＃39;acc&＃39;]
plt.plot(x,y,label&＃61;"acc")y&＃61;history.history[&＃39;loss&＃39;]
plt.plot(x,y,label&＃61;"loss")y&＃61;history.history[&＃39;val_acc&＃39;]
plt.plot(x,y,label&＃61;"val_acc")y&＃61;history.history[&＃39;val_loss&＃39;]
plt.plot(x,y,label&＃61;"val_loss")result &＃61; model.predict(x&＃61;[test_x_1,test_x_2])a&＃61; np.argmax(result,axis&＃61;1)result &＃61; pd.DataFrame()
result[&＃39;Id&＃39;]&＃61;test.id.values
result[&＃39;Category&＃39;]&＃61;[category[i] for i in a]result.to_csv(&＃39;submit.csv&＃39;,index_label&＃61;None)

总体架构图&＃xff1a;

sentence embedding架构图&＃xff1a;

训练呈现&＃xff1a;