NLP采用Bert进行简单文本情感分类
参照当Bert遇上Kerashttps://spaces.ac.cn/archives/6736此示例准确率达到95.5%+
https://github.com/CyberZHG/keras-bert/blob/master/README.zh-CN.md
示例实现
# ! -*- coding:utf-8 -*-
import json
import numpy as np
import pandas as pd
from random import choice
from keras_bert import load_trained_model_from_checkpoint, Tokenizer
import codecs
maxlen = 100
config_path = 'model/bert_config.json'
checkpoint_path = 'model/bert_model.ckpt'
dict_path = 'model/vocab.txt'
token_dict = {}
with codecs.open(dict_path, 'r', 'utf8') as reader:
for line in reader:
token = line.strip()
token_dict[token] = len(token_dict)
class OurTokenizer(Tokenizer):
def __init__(self, token_dict):
super(OurTokenizer, self).__init__(token_dict)
def _tokenize(self, text):
R = []
for c in text:
if c in self._token_dict:
R.append(c)
elif self._is_space(c):
R.append('[unused1]') # space类用未经训练的[unused1]表示
else:
R.append('[UNK]') # 剩余的字符是[UNK]
return R
tokenizer = OurTokenizer(token_dict)
neg = pd.read_excel('neg.xls', header=None)
pos = pd.read_excel('pos.xls', header=None)
data = []
for d in neg[0]:
data.append((d, 0))
for d in pos[0]:
data.append((d, 1))
# 按照9:1的比例划分训练集和验证集
random_order = list(range(len(data)))
np.random.shuffle(random_order)
train_data = [data[j] for i, j in enumerate(random_order) if i % 10 != 0]
valid_data = [data[j] for i, j in enumerate(random_order) if i % 10 == 0]
def seq_padding(X, padding=0):
L = [len(x) for x in X]
ML = max(L)
return np.array([
np.concatenate([x, [padding] * (ML - len(x))]) if len(x) < ML else x for x in X
])
class data_generator:
def __init__(self, data, batch_size=32):
self.data = data
self.batch_size = batch_size
self.steps = len(self.data) // self.batch_size
if len(self.data) % self.batch_size != 0:
self.steps += 1
def __len__(self):
return self.steps
def __iter__(self):
while True:
idxs = list(range(len(self.data)))
np.random.shuffle(idxs)
X1, X2, Y = [], [], []
for i in idxs:
d = self.data[i]
text = d[0][:maxlen]
x1, x2 = tokenizer.encode(first=text)
y = d[1]
X1.append(x1)
X2.append(x2)
Y.append([y])
if len(X1) == self.batch_size or i == idxs[-1]:
X1 = seq_padding(X1)
X2 = seq_padding(X2)
Y = seq_padding(Y)
yield [X1, X2], Y
[X1, X2, Y] = [], [], []
from keras.layers import *
from keras.models import Model
from keras.optimizers import Adam
bert_model = load_trained_model_from_checkpoint(config_path, checkpoint_path, seq_len=None)
for l in bert_model.layers:
l.trainable = False
x1_in = Input(shape=(None,))
x2_in = Input(shape=(None,))
x = bert_model([x1_in, x2_in])
x = Lambda(lambda x: x[:, 0])(x)
p = Dense(1, activation='sigmoid')(x)
model = Model([x1_in, x2_in], p)
model.compile(
loss='binary_crossentropy',
optimizer=Adam(1e-5), # 用足够小的学习率
metrics=['accuracy']
)
model.summary()
train_D = data_generator(train_data)
valid_D = data_generator(valid_data)
test = [train_data[0]]
test_D = data_generator(test)
model.fit_generator(
train_D.__iter__(),
steps_per_epoch=len(train_D),
epochs=1,
validation_data=valid_D.__iter__(),
validation_steps=len(valid_D)
)
#保存模型权重值
model.save('model.h5')
原示例存在的问题
模型在保持完之后再进行加载时提示存在自定义层和激活方法的问题，暂没找到解决办法，如有知道办法的小伙伴请留言私信
问题解决
# ! -*- coding:utf-8 -*-
import json
import numpy as np
import pandas as pd
from random import choice
from keras_bert import load_trained_model_from_checkpoint, Tokenizer, get_custom_objects
import re, os
import codecs
from keras.models import load_model
maxlen = 100
config_path = 'model/bert_config.json'
checkpoint_path = 'model/bert_model.ckpt'
dict_path = 'model/vocab.txt'
token_dict = {}
with codecs.open(dict_path, 'r', 'utf8') as reader:
for line in reader:
token = line.strip()
token_dict[token] = len(token_dict)
class OurTokenizer(Tokenizer):
def __init__(self, token_dict):
super(OurTokenizer, self).__init__(token_dict)
def _tokenize(self, text):
R = []
for c in text:
if c in self._token_dict:
R.append(c)
elif self._is_space(c):
R.append('[unused1]') # space类用未经训练的[unused1]表示
else:
R.append('[UNK]') # 剩余的字符是[UNK]
return R
tokenizer = OurTokenizer(token_dict)
neg = pd.read_excel('neg.xls', header=None)
pos = pd.read_excel('pos.xls', header=None)
data = []
for d in neg[0]:
data.append((d, 0))
for d in pos[0]:
data.append((d, 1))
# 按照9:1的比例划分训练集和验证集
random_order = list(range(len(data)))
np.random.shuffle(random_order)
train_data = [data[j] for i, j in enumerate(random_order) if i % 10 != 0]
valid_data = [data[j] for i, j in enumerate(random_order) if i % 10 == 0]
def seq_padding(X, padding=0):
L = [len(x) for x in X]
ML = max(L)
return np.array([
np.concatenate([x, [padding] * (ML - len(x))]) if len(x) < ML else x for x in X
])
class data_generator:
def __init__(self, data, batch_size=32):
self.data = data
self.batch_size = batch_size
self.steps = len(self.data) // self.batch_size
if len(self.data) % self.batch_size != 0:
self.steps += 1
def __len__(self):
return self.steps
def __iter__(self):
while True:
idxs = list(range(len(self.data)))
np.random.shuffle(idxs)
X1, X2, Y = [], [], []
for i in idxs:
d = self.data[i]
text = d[0][:maxlen]
x1, x2 = tokenizer.encode(first=text)
y = d[1]
X1.append(x1)
X2.append(x2)
Y.append([y])
if len(X1) == self.batch_size or i == idxs[-1]:
X1 = seq_padding(X1)
X2 = seq_padding(X2)
Y = seq_padding(Y)
yield [X1, X2], Y
[X1, X2, Y] = [], [], []
from keras.layers import *
from keras.models import Model
import keras.backend as K
from keras.optimizers import Adam
bert_model = load_trained_model_from_checkpoint(config_path, checkpoint_path, seq_len=None)
for l in bert_model.layers:
l.trainable = False
x1_in = Input(shape=(None,))
x2_in = Input(shape=(None,))
x = bert_model([x1_in, x2_in])
print(bert_model.layers)
x = Lambda(lambda x: x[:, 0])(x)
p = Dense(1, activation='sigmoid')(x)
model = Model([x1_in, x2_in], p)
model.compile(
loss='binary_crossentropy',
optimizer=Adam(1e-5), # 用足够小的学习率
metrics=['accuracy']
)
model.summary()
train_D = data_generator(train_data)
valid_D = data_generator(valid_data)
'''
model.fit_generator(
train_D.__iter__(),
steps_per_epoch=len(train_D),
epochs=5,
validation_data=valid_D.__iter__(),
validation_steps=len(valid_D)
)
model.save('save_path.h5')
'''
# 定义生成器将数据集解析为
class data_token_generator:
def __init__(self, data, batch_size=32):
self.data = data
self.batch_size = batch_size
self.steps = len(self.data) # self.batch_size
if len(self.data) % self.batch_size != 0:
self.steps += 1
def __len__(self):
return self.steps
def get_data(self):
idxs = list(range(len(self.data)))
np.random.shuffle(idxs)
X1, X2, Y = [], [], []
for i in idxs:
d = self.data[i]
text = d[0][:maxlen]
print(text)
x1, x2 = tokenizer.encode(first=text)
y = d[1]
X1.append(x1)
X2.append(x2)
Y.append([y])
X1 = seq_padding(X1)
X2 = seq_padding(X2)
Y = seq_padding(Y)
return X1, X2, Y
new_model = load_model('save_path.h5', custom_objects=get_custom_objects())
test_T = data_token_generator(valid_data[0:10])
X_test1, X_test2, Y_test = test_T.get_data()
print(Y_test)
print(new_model.predict([X_test1, X_test2]))
我的实现
# ! -*- coding:utf-8 -*-
import numpy as np
import pandas as pd
from random import choice
from keras_bert import load_trained_model_from_checkpoint, Tokenizer, get_checkpoint_paths
import codecs
from keras.layers import *
from keras.models import Model
from keras.optimizers import Adam
# 评价文本最大长度
maxlen = 100
dict_path = 'model/vocab.txt'
token_dict = {}
EPOCHS = 30
BATCH_SIZE = 128
# 初始化令牌字典
with codecs.open(dict_path, 'r', 'utf8') as reader:
for line in reader:
token = line.strip()
# print(token, len(token_dict))
token_dict[token] = len(token_dict)
# 定义令牌解析器
class OurTokenizer(Tokenizer):
def _tokenize(self, text):
R = []
for c in text:
if c in self._token_dict:
R.append(c)
elif self._is_space(c):
R.append('[unused1]') # space类用未经训练的[unused1]表示
else:
R.append('[UNK]') # 剩余的字符是[UNK]
return R
# 初始化令牌解析器
tokenizer = OurTokenizer(token_dict)
# 读取数据集
neg = pd.read_excel('neg.xls', header=None)
pos = pd.read_excel('pos.xls', header=None)
data = []
for d in neg[0]:
data.append((d, 0))
for d in pos[0]:
data.append((d, 1))
# 按照9:1的比例划分训练集和验证集
random_order = list(range(len(data)))
np.random.shuffle(random_order)
train_data = [data[j] for i, j in enumerate(random_order) if i % 10 != 0]
valid_data = [data[j] for i, j in enumerate(random_order) if i % 10 == 0]
# 令牌序列长度补全
def seq_padding(X, padding=0):
L = [len(x) for x in X]
ML = max(L)
t = [
np.concatenate([x, [padding] * (ML - len(x))]) if len(x) < ML else x for x in X
]
return t
# 定义生成器将数据集解析为
class data_token_generator:
def __init__(self, data, batch_size=32, print_text=False):
self.data = data
self.batch_size = batch_size
self.steps = len(self.data) # self.batch_size
self.print_text = print_text
if len(self.data) % self.batch_size != 0:
self.steps += 1
# bert中文模型路径
paths = get_checkpoint_paths('model')
# bert中文模型加载
self.bert_model = load_trained_model_from_checkpoint(paths.config, paths.checkpoint, seq_len=None)
for l in self.bert_model.layers:
l.trainable = True
def __len__(self):
return self.steps
def get_data(self):
data_x = []
data_y = []
idxs = list(range(len(self.data)))
# 随机
np.random.shuffle(idxs)
indices, segments, Y = [], [], []
for i in idxs:
d = self.data[i]
# 截取数据
text = d[0][:maxlen]
if self.print_text:
print(text)
# 生成指标及段
indice, segment = tokenizer.encode(first=text)
y = d[1]
# 数据放入数组中
indices.append(indice)
segments.append(segment)
Y.append([y])
# 转化成批次
if len(indices) == self.batch_size or i == idxs[-1]:
indices = seq_padding(indices)
segments = seq_padding(segments)
Y = seq_padding(Y)
# 产生词向量
x = self.bert_model.predict([np.array(indices), np.array(segments)])
j_idxs = list(range(len(x)))
for j in j_idxs:
data_x.append(x[j])
data_y.append(Y[j])
print(len(data_y))
[indices, segments, Y] = [], [], []
return np.array(data_x), np.array(data_y)
# 定义二分类网络
x_in = Input(shape=(None, 768))
x = Lambda(lambda x: x[:, 0])(x_in)
p = Dense(1, activation='sigmoid')(x)
model = Model(x_in, p)
model.compile(
loss='binary_crossentropy',
optimizer=Adam(1e-5), # 用足够小的学习率
metrics=['accuracy']
)
# 打印模型结构
model.summary()
# 开始训练
print('Training -----------')
train_T = data_token_generator(train_data)
train_x, train_y = train_T.get_data()
valid_T = data_token_generator(valid_data)
validation_data = valid_T.get_data()
model.fit(
train_x,
train_y,
epochs=EPOCHS,
batch_size=BATCH_SIZE,
validation_data=validation_data
)
model.save('new_model.h5')
# 加载模型验证
import keras
test_T = data_token_generator(valid_data[0:10], print_text=True)
X_test, Y_test = test_T.get_data()
print(Y_test)
new_model = keras.models.load_model('new_model.h5')
y = new_model.predict(X_test)
print(y)
采用哈工大版权重，准确率在80%左右
相关依赖
- keras_bert https://github.com/CyberZHG/keras-bert
中文版权重
NLP采用Bert进行简单文本情感分类的更多相关文章
- NLP之基于TextCNN的文本情感分类
TextCNN @ 目录 TextCNN 1.理论 1.1 基础概念 最大汇聚(池化)层: 1.2 textCNN模型结构 2.实验 2.1 实验步骤 2.2 算法模型 1.理论 1.1 基础概念 在 ...
- 基于Bert的文本情感分类
详细代码已上传到github: click me Abstract: Sentiment classification is the process of analyzing and reaso ...
- NLP文本情感分类传统模型+深度学习（demo）
文本情感分类: 文本情感分类(一):传统模型 摘自:http://spaces.ac.cn/index.php/archives/3360/ 测试句子:工信处女干事每月经过下属科室都要亲口交代24口交 ...
- NLP之基于Bi-LSTM和注意力机制的文本情感分类
Bi-LSTM(Attention) @ 目录 Bi-LSTM(Attention) 1.理论 1.1 文本分类和预测(翻译) 1.2 注意力模型 1.2.1 Attention模型 1.2.2 Bi ...
- 文本情感分类：分词 OR 不分词(3)
为什么要用深度学习模型?除了它更高精度等原因之外,还有一个重要原因,那就是它是目前唯一的能够实现“端到端”的模型.所谓“端到端”,就是能够直接将原始数据和标签输入,然后让模型自己完成一切过程——包括特 ...
- pytorch 文本情感分类和命名实体识别NER中LSTM输出的区别
文本情感分类: 文本情感分类采用LSTM的最后一层输出 比如双层的LSTM,使用正向的最后一层和反向的最后一层进行拼接 def forward(self,input): ''' :param inpu ...
- kaggle之电影评论文本情感分类
电影文本情感分类 Github地址 Kaggle地址 这个任务主要是对电影评论文本进行情感分类,主要分为正面评论和负面评论,所以是一个二分类问题,二分类模型我们可以选取一些常见的模型比如贝叶斯.逻辑回 ...
- NLP（二十二）利用ALBERT实现文本二分类
在文章NLP(二十)利用BERT实现文本二分类中,笔者介绍了如何使用BERT来实现文本二分类功能,以判别是否属于出访类事件为例子.但是呢,利用BERT在做模型预测的时候存在预测时间较长的问题.因此 ...
- 基于 Spark 的文本情感分析
转载自:https://www.ibm.com/developerworks/cn/cognitive/library/cc-1606-spark-seniment-analysis/index.ht ...
随机推荐
- 反射 go
reflect.Valueof 到底是个什么? 反射值对象(reflect.Value)提供一系列方法进行零值和空判定,如下表所示. 反射值对象的零值和有效性判断方法 方 法 说 明 IsNil() ...
- 学习 C++ 读什么书
C++ 看什么书? 很多莘莘学子,想要在这条路上前进,却是踽踽独行,摸不到门路.就像我当初是一样的.在没有一个前辈带路的情况下,想要自己一个人摸索前进,那份迷茫,深深的刻进自己的记忆里. 下面 ...
- 爬虫 lxml 模块
Xpath 在 XML 文档中查找信息的语言, 同样适用于 HTML 辅助工具 Xpath Helper Chrome插件 快捷键 Ctrl + shift + x XML Quire xpath ...
- linuxs上mono
当前,在Linux系统上架设ASP.NET网站.建设WEB应用工程项目已经在国内流行起来,而“Mono+Jexus”架构模式是Linux承载ASP.NET企业级应用的极为重要的架构方式,这种架构中,J ...
- flutter Could not find the built application bundle at build/ios/iphonesimulator/Runner.app
运行flutter run时报错 提示如下: Could not find the built application bundle at build/ios/iphonesimulator/Runn ...
- Windows下Elasticsearch安装问题处理
按ES官网的安装方法正常安装就行了.可能遇到的其他问题. 1.报jvm.dll不存在. 只需要重新安装JDK过后,会有jdk1.8.0_73目录和jre1.8.0_73目录.因为java就喜欢玩这种“ ...
- flask 学习 (五)
之前照着书去做的时候经常出现一些小问题,由于对于flask核心内容还不甚了解,我觉定先从更简单的做起,再根据别的需要对搭建的网站进行扩展. 上网找了一下,发现这位http://zhanghonglun ...
- 自在因梦 | 威爾伯的Fourth Turning所引發的聯想
2015-05-06 胡因梦 ...
- etcd安全集群三节点扩容至四个节点
规划:先安装三台组建集群,然后扩容一个安全节点进来 .环境: 三台centos7. 主机 192.168.0.91 192.168.0.92 192.168.0.93 都关闭防火墙 都关闭selinu ...
- .net分流抢票助手
官方网站: http://www.12306bypass.com/作者:Cheney.小风分流抢票基于.Net4.0框架开发,在Windows7之后的操作系统可直接打开.其他操作系统如打不开或者打开报 ...