用神经网络训练一个文本分类器

本文由 - 翻译，校稿。未经许可，禁止转载！
英文出处：。欢迎加入。

理解聊天机器人的工作原理是非常重要的。聊天机器人内部一个非常重要的组件就是文本分类器。我们看一下文本分类器的神经网络（ANN）的内部工作原理。

多层神经网络

我们将会使用2层网络（1个隐层）和一个“词包”的方法来组织我们的训练数据。文本分类有3个特点：模式匹配、算法、神经网络。虽然使用多项朴素贝叶斯算法的方法非常有效，但是它有3个致命的缺陷：

这个算法输出一个分数而不是一个概率。我们可以使用概率来忽略特定阈值以下的预测结果。这类似于忽略收音机中的噪声。
这个算法从一个样本中学习一个分类中包含什么，而不是一个分类中不包含什么。一个分类中不包含什么的的学习模式往往也很重要。
不成比例的大训练集的分类将会导致扭曲的分类分数，迫使算法相对于分类规模来调整输出分数，这并不理想。

和它“天真”的对手一样，这种分类器并不试图去理解句子的含义，而仅仅对它进行分类。事实上，所谓的“人工智能聊天机器人”并不理解语言，但那是。

如果你刚接触人工神经网络，这是它的。

理解分类算法，请看。

我们来逐个分析文本分类器的每个部分。我们将按照以下顺序：

引用需要的库
提供训练集
整理数据
迭代：编写代码+测试预测结果+调整模型
抽象

代码在这里，我们使用ipython notebook这个在数据科学项目上非常高效的工具。代码语法是python。

我们首先导入自然语言工具包。我们需要一个可靠的方法将句子切分成词并且将单词词干化处理。

# use natural language toolkit
import nltk
from nltk.stem.lancaster import LancasterStemmer
import os
import json
import datetime
stemmer = LancasterStemmer()

# use natural language toolkit

import nltk

from nltk.stem.lancaster import LancasterStemmer

import os

import json

import datetime

stemmer = LancasterStemmer()

下面是我们的训练集，12个句子属于3个类别（“意图”）。

# 3 classes of training data
training_data = []
training_data.append({"class":"greeting", "sentence":"how are you?"})
training_data.append({"class":"greeting", "sentence":"how is your day?"})
training_data.append({"class":"greeting", "sentence":"good day"})
training_data.append({"class":"greeting", "sentence":"how is it going today?"})

training_data.append({"class":"goodbye", "sentence":"have a nice day"})
training_data.append({"class":"goodbye", "sentence":"see you later"})
training_data.append({"class":"goodbye", "sentence":"have a nice day"})
training_data.append({"class":"goodbye", "sentence":"talk to you soon"})

training_data.append({"class":"sandwich", "sentence":"make me a sandwich"})
training_data.append({"class":"sandwich", "sentence":"can you make a sandwich?"})
training_data.append({"class":"sandwich", "sentence":"having a sandwich today?"})
training_data.append({"class":"sandwich", "sentence":"what's for lunch?"})
print ("%s sentences in training data" % len(training_data))

# 3 classes of training data

training_data = []

training_data.append({"class":"greeting", "sentence":"how are you?"})

training_data.append({"class":"greeting", "sentence":"how is your day?"})

training_data.append({"class":"greeting", "sentence":"good day"})

training_data.append({"class":"greeting", "sentence":"how is it going today?"})

training_data.append({"class":"goodbye", "sentence":"have a nice day"})

training_data.append({"class":"goodbye", "sentence":"see you later"})

training_data.append({"class":"goodbye", "sentence":"have a nice day"})

training_data.append({"class":"goodbye", "sentence":"talk to you soon"})

training_data.append({"class":"sandwich", "sentence":"make me a sandwich"})

training_data.append({"class":"sandwich", "sentence":"can you make a sandwich?"})

training_data.append({"class":"sandwich", "sentence":"having a sandwich today?"})

training_data.append({"class":"sandwich", "sentence":"what's for lunch?"})

print ("%s sentences in training data" % len(training_data))

12 sentences in training data

1	12 sentences in training data

现在我们可以将数据结构组织为：documents, classes 和words.

words = []
classes = []
documents = []
ignore_words = ['?']
# loop through each sentence in our training data
for pattern in training_data:
    # tokenize each word in the sentence
    w = nltk.word_tokenize(pattern['sentence'])
    # add to our words list
    words.extend(w)
    # add to documents in our corpus
    documents.append((w, pattern['class']))
    # add to our classes list
    if pattern['class'] not in classes:
        classes.append(pattern['class'])

# stem and lower each word and remove duplicates
words = [stemmer.stem(w.lower()) for w in words if w not in ignore_words]
words = list(set(words))

# remove duplicates
classes = list(set(classes))

print (len(documents), "documents")
print (len(classes), "classes", classes)
print (len(words), "unique stemmed words", words)

words = []

classes = []

documents = []

ignore_words = ['?']

# loop through each sentence in our training data

for pattern in training_data:

# tokenize each word in the sentence

w = nltk.word_tokenize(pattern['sentence'])

# add to our words list

words.extend(w)

# add to documents in our corpus

documents.append((w, pattern['class']))

# add to our classes list

if pattern['class'] not in classes:

classes.append(pattern['class'])

# stem and lower each word and remove duplicates

words = [stemmer.stem(w.lower()) for w in words if w not in ignore_words]

words = list(set(words))

# remove duplicates

classes = list(set(classes))

print (len(documents), "documents")

print (len(classes), "classes", classes)

print (len(words), "unique stemmed words", words)

12 documents
3 classes ['greeting', 'goodbye', 'sandwich']
26 unique stemmed words ['sandwich', 'hav', 'a', 'how', 'for', 'ar', 'good', 'mak', 'me', 'it', 'day', 'soon', 'nic', 'lat', 'going', 'you', 'today', 'can', 'lunch', 'is', "'s", 'see', 'to', 'talk', 'yo', 'what']

12 documents

3 classes ['greeting', 'goodbye', 'sandwich']

26 unique stemmed words ['sandwich', 'hav', 'a', 'how', 'for', 'ar', 'good', 'mak', 'me', 'it', 'day', 'soon', 'nic', 'lat', 'going', 'you', 'today', 'can', 'lunch', 'is', "'s", 'see', 'to', 'talk', 'yo', 'what']

注意每个单词都是词根并且小写。词根有助于机器将“have”和“having”等同起来。同时我们也不关心大小写。

我们将训练集中的每个句子转换为词包。

# create our training data
training = []
output = []
# create an empty array for our output
output_empty = [0] * len(classes)

# training set, bag of words for each sentence
for doc in documents:
    # initialize our bag of words
    bag = []
    # list of tokenized words for the pattern
    pattern_words = doc[0]
    # stem each word
    pattern_words = [stemmer.stem(word.lower()) for word in pattern_words]
    # create our bag of words array
    for w in words:
        bag.append(1) if w in pattern_words else bag.append(0)

    training.append(bag)
    # output is a '0' for each tag and '1' for current tag
    output_row = list(output_empty)
    output_row[classes.index(doc[1])] = 1
    output.append(output_row)

# sample training/output
i = 0
w = documents[i][0]
print ([stemmer.stem(word.lower()) for word in w])
print (training[i])
print (output[i])

# create our training data

training = []

output = []

# create an empty array for our output

output_empty = [0] * len(classes)

# training set, bag of words for each sentence

for doc in documents:

# initialize our bag of words

bag = []

# list of tokenized words for the pattern

pattern_words = doc[0]

# stem each word

pattern_words = [stemmer.stem(word.lower()) for word in pattern_words]

# create our bag of words array

for w in words:

bag.append(1) if w in pattern_words else bag.append(0)

training.append(bag)

# output is a '0' for each tag and '1' for current tag

output_row = list(output_empty)

output_row[classes.index(doc[1])] = 1

output.append(output_row)

# sample training/output

i = 0

w = documents[i][0]

print ([stemmer.stem(word.lower()) for word in w])

print (training[i])

print (output[i])

['how', 'ar', 'you', '?']
[0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
[1, 0, 0]

['how', 'ar', 'you', '?']

[0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

[1, 0, 0]

上面的步骤是文本分类中的一个经典步骤：每个训练句子被转化为一个包含0和1的数组，而不是语料库中包含独特单词的数组。

['how', 'are', 'you', '?']

1	['how', 'are', 'you', '?']

被词干化为：

['how', 'ar', 'you', '?']

1	['how', 'ar', 'you', '?']

然后转换为输入词包的形式：1代表单词存在于词包中（忽略问号？）

[0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

1	[0, 0, 0, 1, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]

输出：第一类

[1, 0, 0]

[1, 0, 0]

注意：一个句子可以有多个分类，也可以没有。确保理解上面的内容，仔细阅读代码直到你理解它。

机器学习的第一步是要有干净的数据

接下来我们的学习2层神经网络的核心功能。

如果你是人工神经网络新手，这里是它的工作原理

我们使用，原因是它可以提供快速的矩阵乘法运算。

我们使用sigmoid函数对值进行归一化，用其导数来衡量错误率。通过不断迭代和调整，直到错误率低到一个可以接受的值。

下面我们也实现了bag-of-words函数，将输入的一个句子转化为一个包含0和1的数组。这就是转换训练数据，得到正确的转换数据至关重要。

import numpy as np
import time

# compute sigmoid nonlinearity
def sigmoid(x):
    output = 1/(1+np.exp(-x))
    return output

# convert output of sigmoid function to its derivative
def sigmoid_output_to_derivative(output):
    return output*(1-output)

def clean_up_sentence(sentence):
    # tokenize the pattern
    sentence_words = nltk.word_tokenize(sentence)
    # stem each word
    sentence_words = [stemmer.stem(word.lower()) for word in sentence_words]
    return sentence_words

# return bag of words array: 0 or 1 for each word in the bag that exists in the sentence
def bow(sentence, words, show_details=False):
    # tokenize the pattern
    sentence_words = clean_up_sentence(sentence)
    # bag of words
    bag = [0]*len(words)  
    for s in sentence_words:
        for i,w in enumerate(words):
            if w == s: 
                bag[i] = 1
                if show_details:
                    print ("found in bag: %s" % w)

    return(np.array(bag))

def think(sentence, show_details=False):
    x = bow(sentence.lower(), words, show_details)
    if show_details:
        print ("sentence:", sentence, "n bow:", x)
    # input layer is our bag of words
    l0 = x
    # matrix multiplication of input and hidden layer
    l1 = sigmoid(np.dot(l0, synapse_0))
    # output layer
    l2 = sigmoid(np.dot(l1, synapse_1))
    return l2

import numpy as np

import time

# compute sigmoid nonlinearity

def sigmoid(x):

output = 1/(1+np.exp(-x))

return output

# convert output of sigmoid function to its derivative

def sigmoid_output_to_derivative(output):

return output*(1-output)

def clean_up_sentence(sentence):

# tokenize the pattern

sentence_words = nltk.word_tokenize(sentence)

# stem each word

sentence_words = [stemmer.stem(word.lower()) for word in sentence_words]

return sentence_words

# return bag of words array: 0 or 1 for each word in the bag that exists in the sentence

def bow(sentence, words, show_details=False):

# tokenize the pattern

sentence_words = clean_up_sentence(sentence)

# bag of words

bag = [0]*len(words)

for s in sentence_words:

for i,w in enumerate(words):

if w == s:

bag[i] = 1

if show_details:

print ("found in bag: %s" % w)

return(np.array(bag))

def think(sentence, show_details=False):

x = bow(sentence.lower(), words, show_details)

if show_details:

print ("sentence:", sentence, "n bow:", x)

# input layer is our bag of words

l0 = x

# matrix multiplication of input and hidden layer

l1 = sigmoid(np.dot(l0, synapse_0))

# output layer

l2 = sigmoid(np.dot(l1, synapse_1))

return l2

现在我们对神经网络训练函数进行编码，创造连接权重。别太激动，这主要是矩阵乘法——来自中学数学课堂。

我们现在准备去构建我们的神经网络模型，我们将连接权重保存为json文件。

你应该尝试不同的“α”（梯度下降参数），看看它是如何影响错误率。此参数有助于错误调整，并找到最低错误率：

synapse_0 += alpha * synapse_0_weight_update

我们在隐藏层使用了20个神经元，你可以很容易地调整。这些参数将随着于您的训练数据规模的不同而不同，将错误率调整到低于10 ^ – 3是比较合理的。

X = np.array(training)
y = np.array(output)

start_time = time.time()

train(X, y, hidden_neurons=20, alpha=0.1, epochs=100000, dropout=False, dropout_percent=0.2)

elapsed_time = time.time() - start_time
print ("processing time:", elapsed_time, "seconds")

X = np.array(training)

y = np.array(output)

start_time = time.time()

train(X, y, hidden_neurons=20, alpha=0.1, epochs=100000, dropout=False, dropout_percent=0.2)

elapsed_time = time.time() - start_time

print ("processing time:", elapsed_time, "seconds")

Training with 20 neurons, alpha:0.1, dropout:False 
Input matrix: 12x26    Output matrix: 1x3
delta after 10000 iterations:0.0062613597435
delta after 20000 iterations:0.00428296074919
delta after 30000 iterations:0.00343930779307
delta after 40000 iterations:0.00294648034566
delta after 50000 iterations:0.00261467859609
delta after 60000 iterations:0.00237219554105
delta after 70000 iterations:0.00218521899378
delta after 80000 iterations:0.00203547284581
delta after 90000 iterations:0.00191211022401
delta after 100000 iterations:0.00180823798397
saved synapses to: synapses.json
processing time: 6.501226902008057 seconds

Training with 20 neurons, alpha:0.1, dropout:False

Input matrix: 12x26 Output matrix: 1x3

delta after 10000 iterations:0.0062613597435

delta after 20000 iterations:0.00428296074919

delta after 30000 iterations:0.00343930779307

delta after 40000 iterations:0.00294648034566

delta after 50000 iterations:0.00261467859609

delta after 60000 iterations:0.00237219554105

delta after 70000 iterations:0.00218521899378

delta after 80000 iterations:0.00203547284581

delta after 90000 iterations:0.00191211022401

delta after 100000 iterations:0.00180823798397

saved synapses to: synapses.json

processing time: 6.501226902008057 seconds

synapse.json文件中包含了全部的连接权重，这就是我们的模型。

一旦连接权重已经计算完成，对于分类来说只需要classify()函数了：大约15行代码

备注：如果训练集有变化，我们的模型需要重新计算。对于非常大的数据集，这需要较长的时间。

现在我们可以生成一个句子属于一个或者多个分类的概率了。它的速度非常快，这是因为我们之前定义的think()函数中的点积运算。

# probability threshold
ERROR_THRESHOLD = 0.2
# load our calculated synapse values
synapse_file = 'synapses.json' 
with open(synapse_file) as data_file: 
    synapse = json.load(data_file) 
    synapse_0 = np.asarray(synapse['synapse0']) 
    synapse_1 = np.asarray(synapse['synapse1'])

def classify(sentence, show_details=False):
    results = think(sentence, show_details)

    results = [[i,r] for i,r in enumerate(results) if r>ERROR_THRESHOLD ] 
    results.sort(key=lambda x: x[1], reverse=True) 
    return_results =[[classes[r[0]],r[1]] for r in results]
    print ("%s n classification: %s" % (sentence, return_results))
    return return_results

classify("sudo make me a sandwich")
classify("how are you today?")
classify("talk to you tomorrow")
classify("who are you?")
classify("make me some lunch")
classify("how was your lunch today?")
print()
classify("good day", show_details=True)

# probability threshold

ERROR_THRESHOLD = 0.2

# load our calculated synapse values

synapse_file = 'synapses.json'

with open(synapse_file) as data_file:

synapse = json.load(data_file)

synapse_0 = np.asarray(synapse['synapse0'])

synapse_1 = np.asarray(synapse['synapse1'])

def classify(sentence, show_details=False):

results = think(sentence, show_details)

results = [[i,r] for i,r in enumerate(results) if r>ERROR_THRESHOLD ]

results.sort(key=lambda x: x[1], reverse=True)

return_results =[[classes[r[0]],r[1]] for r in results]

print ("%s n classification: %s" % (sentence, return_results))

return return_results

classify("sudo make me a sandwich")

classify("how are you today?")

classify("talk to you tomorrow")

classify("who are you?")

classify("make me some lunch")

classify("how was your lunch today?")

print()

classify("good day", show_details=True)

<strong>sudo make me a sandwich </strong>
 [['sandwich', 0.99917711814437993]]
<strong>how are you today? </strong>
 [['greeting', 0.99864563257858363]]
<strong>talk to you tomorrow </strong>
 [['goodbye', 0.95647479275905511]]
<strong>who are you? </strong>
 [['greeting', 0.8964283843977312]]
<strong>make me some lunch</strong> 
 [['sandwich', 0.95371924052636048]]
<strong>how was your lunch today? </strong>
 [['greeting', 0.99120883810944971], ['sandwich', 0.31626066870883057]]

sudo make me a sandwich

[['sandwich', 0.99917711814437993]]

how are you today?

[['greeting', 0.99864563257858363]]

talk to you tomorrow

[['goodbye', 0.95647479275905511]]

who are you?

[['greeting', 0.8964283843977312]]

make me some lunch

[['sandwich', 0.95371924052636048]]

how was your lunch today?

[['greeting', 0.99120883810944971], ['sandwich', 0.31626066870883057]]

你可以用其它语句、不同概率来试验几次，也可以添加训练数据来改进／扩展当前的模型。尤其注意用很少的训练数据就得到稳定的预测结果。

有一些句子将会产生多个预测结果（高于阈值）。你需要给你的程序设定一个合适的阈值。并非所有的文本分类方案都是相同的：一些预测情况比其他预测需要更高的置信水平。

最后这个分类结果展示了一些内部的细节：

found in bag: good
found in bag: day
sentence: **good day** 
 bow: [0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
good day 
 [['greeting', 0.99664077655648697]]

found in bag: good

found in bag: day

sentence: **good day**

bow: [0 0 0 0 0 0 1 0 0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]

good day

[['greeting', 0.99664077655648697]]

从这个句子的词包中可以看到，有两个单词和我们的词库是匹配的。同时我们的神经网络从这些 0 代表的非匹配词语中学习了。

如果提供一个仅仅有一个常用单词 ‘a’ 被匹配的句子，那我们会得到一个低概率的分类结果A：

found in bag: a
sentence: **a burrito! **
 bow: [0 0 1 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
a burrito! 
 [['sandwich', 0.61776860634647834]]