我们有一个数据集，其中包含 Twitter 上的推文及其对网络欺凌攻击类型的分类。 我们需要使用 NLP 来预处理数据并构建一个 ML 模型，该模型可以在这些数据集上执行以高精度对目标变量进行分类。
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import seaborn as sns
df = pd.read_csv("cyberbullying_tweets.csv")
df.head()
| tweet_text | cyberbullying_type | |
|---|---|---|
| 0 | In other words #katandandre, your food was cra... | not_cyberbullying |
| 1 | Why is #aussietv so white? #MKR #theblock #ImA... | not_cyberbullying |
| 2 | @XochitlSuckkks a classy whore? Or more red ve... | not_cyberbullying |
| 3 | @Jason_Gio meh. :P thanks for the heads up, b... | not_cyberbullying |
| 4 | @RudhoeEnglish This is an ISIS account pretend... | not_cyberbullying |
df['tweet_text'].iloc[47000]
'Call me disgusting then! Because if any non BLACK person, especially a white one, calls me a nigger I’m going to slap the dogshit out of them too...it’s a new day for the Karens around the world.'
df.shape
(47692, 2)
# 检查空值
df.isna().sum()
tweet_text 0 cyberbullying_type 0 dtype: int64
# 检查重复值
df.duplicated().sum()
36
df = df[~df.duplicated()]
df.info()
<class 'pandas.core.frame.DataFrame'> Int64Index: 47656 entries, 0 to 47691 Data columns (total 2 columns): # Column Non-Null Count Dtype --- ------ -------------- ----- 0 tweet_text 47656 non-null object 1 cyberbullying_type 47656 non-null object dtypes: object(2) memory usage: 1.1+ MB
df['cyberbullying_type'].value_counts()
religion 7997 age 7992 ethnicity 7959 gender 7948 not_cyberbullying 7937 other_cyberbullying 7823 Name: cyberbullying_type, dtype: int64
# 文本标签有五种，展示其分布
sns.displot(data=df,x='cyberbullying_type')
plt.xticks(rotation=90)
([0, 1, 2, 3, 4, 5], [Text(0, 0, 'not_cyberbullying'), Text(1, 0, 'gender'), Text(2, 0, 'religion'), Text(3, 0, 'other_cyberbullying'), Text(4, 0, 'age'), Text(5, 0, 'ethnicity')])
分布显示数据集非常均衡，这是一个很好的事情。
def remove_punctuation(text):
'''删除标点符号的函数'''
import string
# 将标点符号替换为无空格，
# 这实际上删除了标点符号
translator = str.maketrans('', '', string.punctuation)
# 返回去掉标点符号的文本
return text.translate(translator)
# 查看标点符号
import string
string.punctuation
'!"#$%&\'()*+,-./:;<=>?@[\\]^_`{|}~'
# 应用函数删除标点符号
df['tweet_text'] = df['tweet_text'].apply(remove_punctuation)
df.head(10)
| tweet_text | cyberbullying_type | |
|---|---|---|
| 0 | In other words katandandre your food was crapi... | not_cyberbullying |
| 1 | Why is aussietv so white MKR theblock ImACeleb... | not_cyberbullying |
| 2 | XochitlSuckkks a classy whore Or more red velv... | not_cyberbullying |
| 3 | JasonGio meh P thanks for the heads up but no... | not_cyberbullying |
| 4 | RudhoeEnglish This is an ISIS account pretendi... | not_cyberbullying |
| 5 | Raja5aab Quickieleaks Yes the test of god is t... | not_cyberbullying |
| 6 | Itu sekolah ya bukan tempat bully Ga jauh kaya... | not_cyberbullying |
| 7 | Karma I hope it bites Kat on the butt She is j... | not_cyberbullying |
| 8 | stockputout everything but mostly my priest | not_cyberbullying |
| 9 | Rebecca Black Drops Out of School Due to Bullying | not_cyberbullying |
df['tweet_text'].iloc[29966]
'SERIOUSLY not ANOTHER instant restaurant round 😡 MKR'
# 从文本中清除表情符号的功能
import re
def remove_emojis(text):
emoj = re.compile("["
u"\U0001F600-\U0001F64F" # emoticons
u"\U0001F300-\U0001F5FF" # symbols & pictographs
u"\U0001F680-\U0001F6FF" # transport & map symbols
u"\U0001F1E0-\U0001F1FF" # flags (iOS)
u"\U00002500-\U00002BEF" # chinese char
u"\U00002702-\U000027B0"
u"\U00002702-\U000027B0"
u"\U000024C2-\U0001F251"
u"\U0001f926-\U0001f937"
u"\U00010000-\U0010ffff"
u"\u2640-\u2642"
u"\u2600-\u2B55"
u"\u200d"
u"\u23cf"
u"\u23e9"
u"\u231a"
u"\ufe0f" # dingbats
u"\u3030"
"]+", re.UNICODE)
return re.sub(emoj, '', text)
# 应用该功能清除表情符号
df['tweet_text'] = df['tweet_text'].apply(remove_emojis)
df['tweet_text'].iloc[29966]
'SERIOUSLY not ANOTHER instant restaurant round MKR'
提取停用词
# 下载停用词
import nltk
nltk.download('stopwords')
[nltk_data] Downloading package stopwords to [nltk_data] C:\Users\1\AppData\Roaming\nltk_data... [nltk_data] Package stopwords is already up-to-date!
True
# 从nltk library中提取停用词
from nltk.corpus import stopwords
sw = stopwords.words('english')
# 展示停用词
np.array(sw)
array(['i', 'me', 'my', 'myself', 'we', 'our', 'ours', 'ourselves', 'you',
"you're", "you've", "you'll", "you'd", 'your', 'yours', 'yourself',
'yourselves', 'he', 'him', 'his', 'himself', 'she', "she's", 'her',
'hers', 'herself', 'it', "it's", 'its', 'itself', 'they', 'them',
'their', 'theirs', 'themselves', 'what', 'which', 'who', 'whom',
'this', 'that', "that'll", 'these', 'those', 'am', 'is', 'are',
'was', 'were', 'be', 'been', 'being', 'have', 'has', 'had',
'having', 'do', 'does', 'did', 'doing', 'a', 'an', 'the', 'and',
'but', 'if', 'or', 'because', 'as', 'until', 'while', 'of', 'at',
'by', 'for', 'with', 'about', 'against', 'between', 'into',
'through', 'during', 'before', 'after', 'above', 'below', 'to',
'from', 'up', 'down', 'in', 'out', 'on', 'off', 'over', 'under',
'again', 'further', 'then', 'once', 'here', 'there', 'when',
'where', 'why', 'how', 'all', 'any', 'both', 'each', 'few', 'more',
'most', 'other', 'some', 'such', 'no', 'nor', 'not', 'only', 'own',
'same', 'so', 'than', 'too', 'very', 's', 't', 'can', 'will',
'just', 'don', "don't", 'should', "should've", 'now', 'd', 'll',
'm', 'o', 're', 've', 'y', 'ain', 'aren', "aren't", 'couldn',
"couldn't", 'didn', "didn't", 'doesn', "doesn't", 'hadn', "hadn't",
'hasn', "hasn't", 'haven', "haven't", 'isn', "isn't", 'ma',
'mightn', "mightn't", 'mustn', "mustn't", 'needn', "needn't",
'shan', "shan't", 'shouldn', "shouldn't", 'wasn', "wasn't",
'weren', "weren't", 'won', "won't", 'wouldn', "wouldn't"],
dtype='<U10')
elem_remove = ['no', 'nor', 'not',"aren't", 'couldn',
"couldn't", 'didn', "didn't", 'doesn', "doesn't", 'hadn', "hadn't",
'hasn', "hasn't", 'haven', "haven't", 'isn', "isn't", 'ma',
'mightn', "mightn't", 'mustn', "mustn't", 'needn', "needn't",
'shan', "shan't", 'shouldn', "shouldn't", 'wasn', "wasn't",
'weren', "weren't", 'won', "won't", 'wouldn', "wouldn't"]
for elem in elem_remove:
sw.remove(elem)
np.array(sw)
array(['i', 'me', 'my', 'myself', 'we', 'our', 'ours', 'ourselves', 'you',
"you're", "you've", "you'll", "you'd", 'your', 'yours', 'yourself',
'yourselves', 'he', 'him', 'his', 'himself', 'she', "she's", 'her',
'hers', 'herself', 'it', "it's", 'its', 'itself', 'they', 'them',
'their', 'theirs', 'themselves', 'what', 'which', 'who', 'whom',
'this', 'that', "that'll", 'these', 'those', 'am', 'is', 'are',
'was', 'were', 'be', 'been', 'being', 'have', 'has', 'had',
'having', 'do', 'does', 'did', 'doing', 'a', 'an', 'the', 'and',
'but', 'if', 'or', 'because', 'as', 'until', 'while', 'of', 'at',
'by', 'for', 'with', 'about', 'against', 'between', 'into',
'through', 'during', 'before', 'after', 'above', 'below', 'to',
'from', 'up', 'down', 'in', 'out', 'on', 'off', 'over', 'under',
'again', 'further', 'then', 'once', 'here', 'there', 'when',
'where', 'why', 'how', 'all', 'any', 'both', 'each', 'few', 'more',
'most', 'other', 'some', 'such', 'only', 'own', 'same', 'so',
'than', 'too', 'very', 's', 't', 'can', 'will', 'just', 'don',
"don't", 'should', "should've", 'now', 'd', 'll', 'm', 'o', 're',
've', 'y', 'ain', 'aren'], dtype='<U10')
def stopwords(text):
'''a function for removing the stopword'''
# 删除停用词并小写所选单词
text = [word.lower() for word in text.split() if word.lower() not in sw]
# 用空格分隔符连接单词列表
return " ".join(text)
# 应用函数删除停用词
df['tweet_text'] = df['tweet_text'].apply(stopwords)
df.head(10)
| tweet_text | cyberbullying_type | |
|---|---|---|
| 0 | words katandandre food crapilicious mkr | not_cyberbullying |
| 1 | aussietv white mkr theblock imacelebrityau tod... | not_cyberbullying |
| 2 | xochitlsuckkks classy whore red velvet cupcakes | not_cyberbullying |
| 3 | jasongio meh p thanks heads not concerned anot... | not_cyberbullying |
| 4 | rudhoeenglish isis account pretending kurdish ... | not_cyberbullying |
| 5 | raja5aab quickieleaks yes test god good bad in... | not_cyberbullying |
| 6 | itu sekolah ya bukan tempat bully ga jauh kaya... | not_cyberbullying |
| 7 | karma hope bites kat butt nasty mkr | not_cyberbullying |
| 8 | stockputout everything mostly priest | not_cyberbullying |
| 9 | rebecca black drops school due bullying | not_cyberbullying |
df['tweet_text'].iloc[29966]
'seriously not another instant restaurant round mkr'
# 创建清洗函数使其更干净
def clean(text):
text = text.replace('\r', '').replace('\n', ' ')
text = re.sub(r"(?:\@|https?\://)\S+", "", text)
text = re.sub(r'[^\x00-\x7f]',r'', text)
text = re.sub(r"((www.[^s]+)|(http\S+))","",text)
text = re.sub('[0-9]+', '', text)
text = [word for word in text.split() if len(word)< 14 ] # 删除超过 14 个字符的单词
text = " ".join(text)
return text
# 应用清洗功能
df['tweet_text'] = df['tweet_text'].apply(clean)
df.head(10)
| tweet_text | cyberbullying_type | |
|---|---|---|
| 0 | words katandandre food crapilicious mkr | not_cyberbullying |
| 1 | aussietv white mkr theblock today sunrise stud... | not_cyberbullying |
| 2 | classy whore red velvet cupcakes | not_cyberbullying |
| 3 | jasongio meh p thanks heads not concerned anot... | not_cyberbullying |
| 4 | rudhoeenglish isis account pretending kurdish ... | not_cyberbullying |
| 5 | rajaaab quickieleaks yes test god good bad ind... | not_cyberbullying |
| 6 | itu sekolah ya bukan tempat bully ga jauh kaya... | not_cyberbullying |
| 7 | karma hope bites kat butt nasty mkr | not_cyberbullying |
| 8 | stockputout everything mostly priest | not_cyberbullying |
| 9 | rebecca black drops school due bullying | not_cyberbullying |
# 使用正则表达式来识别和替换常见的缩写形式。
def decontract(text):
text = re.sub(r"can\'t", "can not", text)
text = re.sub(r"n\'t", " not", text)
text = re.sub(r"\'re", " are", text)
text = re.sub(r"\'s", " is", text)
text = re.sub(r"\'d", " would", text)
text = re.sub(r"\'ll", " will", text)
text = re.sub(r"\'t", " not", text)
text = re.sub(r"\'ve", " have", text)
text = re.sub(r"\'m", " am", text)
return text
df['tweet_text'] = df['tweet_text'].apply(decontract)
# 清洗和去除重复字符
def clean_repeating_characters(text):
return re.sub(r'(.)1+', r'1', text)
df['tweet_text'] = df['tweet_text'].apply(lambda x: clean_repeating_characters(x))
from nltk.tokenize import RegexpTokenizer
tokenizer = RegexpTokenizer('\w+')
df['tweet_text'] = df['tweet_text'].apply(tokenizer.tokenize)
df['tweet_text'][100]
['mirebotan', 'ramiallolah', 'support', 'dont', 'like', 'people', 'like', 'rami', 'complain', 'things', 'supports', 'done', 'others']
import nltk
st = nltk.PorterStemmer()
def text_stemming(text):
text = [st.stem(word) for word in text]
return text
df['tweet_text'] = df['tweet_text'].apply(lambda x: text_stemming(x))
df['tweet_text'][100]
['mirebotan', 'ramiallolah', 'support', 'dont', 'like', 'peopl', 'like', 'rami', 'complain', 'thing', 'support', 'done', 'other']
# lemmatization
nltk.download('wordnet')
nltk.download('omw-1.4')
lm = nltk.WordNetLemmatizer()
def text_lemmatization(text):
text = [lm.lemmatize(word) for word in text]
return text
df['tweet_text'] = df['tweet_text'].apply(lambda x: text_lemmatization(x))
[nltk_data] Downloading package wordnet to [nltk_data] C:\Users\1\AppData\Roaming\nltk_data... [nltk_data] Downloading package omw-1.4 to [nltk_data] C:\Users\1\AppData\Roaming\nltk_data...
df['tweet_text'].head()
0 [word, katandandr, food, crapilici, mkr] 1 [aussietv, white, mkr, theblock, today, sunris... 2 [classi, whore, red, velvet, cupcak] 3 [jasongio, meh, p, thank, head, not, concern, ... 4 [rudhoeenglish, isi, account, pretend, kurdish... Name: tweet_text, dtype: object
# 创建不包括 not_cyberbullying 的数据框
new_df = df.copy()
new_df = new_df[new_df['cyberbullying_type'] != 'not_cyberbullying']
word_df = new_df['tweet_text'].apply(lambda x: " ".join(x))
word_df.head()
7945 rape realzvasiyana nema joke drunk gay lesbian... 7946 never saw celebr say anyth like obama b maher ... 7947 mean he gay use gender slur make rape joke 7948 rt raulnovoa alexaim mecaesm feminazi 7949 rape rape fact read one post guy get rape comm... Name: tweet_text, dtype: object
# 画词云
from wordcloud import WordCloud
plt.figure(figsize= (20,20))
wc = WordCloud(max_words=1000, width= 1600, height= 800,
collocations= False).generate(' '.join(word_df))
plt.imshow(wc)
<matplotlib.image.AxesImage at 0x1686a708ed0>
# 原数据
df['tweet_text'] = df['tweet_text'].apply(lambda x : " ".join(x))
df.head()
| tweet_text | cyberbullying_type | |
|---|---|---|
| 0 | word katandandr food crapilici mkr | not_cyberbullying |
| 1 | aussietv white mkr theblock today sunris studi... | not_cyberbullying |
| 2 | classi whore red velvet cupcak | not_cyberbullying |
| 3 | jasongio meh p thank head not concern anoth an... | not_cyberbullying |
| 4 | rudhoeenglish isi account pretend kurdish acco... | not_cyberbullying |
再次检查是否有重复条目
df.duplicated().sum()
780
df = df[~df.duplicated()]
df.duplicated().sum()
0
df['cyberbullying_type'].value_counts()
religion 7953 age 7895 ethnicity 7834 not_cyberbullying 7803 gender 7740 other_cyberbullying 7651 Name: cyberbullying_type, dtype: int64
# 创建情绪特征（cyberbullying_type 的副本）
df['sentiment'] = df['cyberbullying_type']
# 将文本的真值对应成标量
df['cyberbullying_type'] = df['cyberbullying_type'].replace({'not_cyberbullying':0,'religion':1,'age':2,'gender':3,'ethnicity':4,'other_cyberbullying':5})
df.head()
| tweet_text | cyberbullying_type | sentiment | |
|---|---|---|---|
| 0 | word katandandr food crapilici mkr | 0 | not_cyberbullying |
| 1 | aussietv white mkr theblock today sunris studi... | 0 | not_cyberbullying |
| 2 | classi whore red velvet cupcak | 0 | not_cyberbullying |
| 3 | jasongio meh p thank head not concern anoth an... | 0 | not_cyberbullying |
| 4 | rudhoeenglish isi account pretend kurdish acco... | 0 | not_cyberbullying |
df['cyberbullying_type'].value_counts()
1 7953 2 7895 4 7834 0 7803 3 7740 5 7651 Name: cyberbullying_type, dtype: int64
# applying countvectorizer and TfidfTransformer
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfTransformer
tfidf = TfidfTransformer()
clf = CountVectorizer()
X_cv = clf.fit_transform(df['tweet_text'])
tf_transformer = TfidfTransformer(use_idf=True).fit(X_cv)
X_tf = tf_transformer.transform(X_cv)
X_tf
<46876x44700 sparse matrix of type '<class 'numpy.float64'>' with 602387 stored elements in Compressed Sparse Row format>
# train and test split
from sklearn.model_selection import train_test_split
X_train, X_test, y_train, y_test = train_test_split(X_tf, df['cyberbullying_type'], test_size=0.20, stratify=df['cyberbullying_type'], random_state=10)
对于多类分类，最适合的 ML 模型是：
1. 朴素贝叶斯
2. K 最近邻
3. 随机森林
我们将所有模型一一拟合并比较结果以获得最佳模型。
#Fitting the model
from sklearn.naive_bayes import MultinomialNB
nb_clf = MultinomialNB()
nb_clf.fit(X_train, y_train)
MultinomialNB()In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
MultinomialNB()
#Evaluating the model
print("Evaluation of Naive Bayes model\n\n")
y_pred_nb = nb_clf.predict(X_test)
from sklearn.metrics import classification_report, confusion_matrix, accuracy_score
result = confusion_matrix(y_test, y_pred_nb)
print(f"Confusion Matrix:\n{result}")
result1 = classification_report(y_test, y_pred_nb)
print(f"Classification Report:\n{result1}")
result2 = accuracy_score(y_test,y_pred_nb)
print(f"Accuracy:{round(result2,2)}\n\n")
Evaluation of Naive Bayes model
Confusion Matrix:
[[ 512 206 315 119 70 339]
[ 8 1555 11 9 7 1]
[ 1 11 1559 2 5 1]
[ 72 36 48 1332 32 28]
[ 3 55 88 10 1403 8]
[ 151 105 263 153 148 710]]
Classification Report:
precision recall f1-score support
0 0.69 0.33 0.44 1561
1 0.79 0.98 0.87 1591
2 0.68 0.99 0.81 1579
3 0.82 0.86 0.84 1548
4 0.84 0.90 0.87 1567
5 0.65 0.46 0.54 1530
accuracy 0.75 9376
macro avg 0.75 0.75 0.73 9376
weighted avg 0.75 0.75 0.73 9376
Accuracy:0.75
#Fitting the model
from sklearn.neighbors import KNeighborsClassifier
knn_clf = KNeighborsClassifier(n_neighbors=5)
knn_clf.fit(X_train, y_train)
KNeighborsClassifier()In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
KNeighborsClassifier()
#Evaluating the model
print("Evaluation of K Nearest Neighbor model\n\n")
y_pred_knn = knn_clf.predict(X_test)
from sklearn.metrics import classification_report, confusion_matrix, accuracy_score
result = confusion_matrix(y_test, y_pred_knn)
print(f"Confusion Matrix:\n{result}")
result1 = classification_report(y_test, y_pred_knn)
print(f"Classification Report:\n{result1}")
result2 = accuracy_score(y_test,y_pred_knn)
print(f"Accuracy:{round(result2,2)}\n\n")
Evaluation of K Nearest Neighbor model
Confusion Matrix:
[[1469 10 3 19 1 59]
[1502 86 0 1 0 2]
[1434 0 139 0 1 5]
[1329 0 1 212 3 3]
[1243 0 1 0 317 6]
[1435 0 3 21 3 68]]
Classification Report:
precision recall f1-score support
0 0.17 0.94 0.29 1561
1 0.90 0.05 0.10 1591
2 0.95 0.09 0.16 1579
3 0.84 0.14 0.24 1548
4 0.98 0.20 0.34 1567
5 0.48 0.04 0.08 1530
accuracy 0.24 9376
macro avg 0.72 0.24 0.20 9376
weighted avg 0.72 0.24 0.20 9376
Accuracy:0.24
0.24的准确度代表K最邻近算法的训练是失败的，因为目标类别只有五类，0.24的准确度还赶不上瞎猜。
#Fitting the model
from sklearn.ensemble import RandomForestClassifier
rf_clf = RandomForestClassifier()
rf_clf.fit(X_train, y_train)
RandomForestClassifier()In a Jupyter environment, please rerun this cell to show the HTML representation or trust the notebook.
RandomForestClassifier()
#Evaluating the model
print("Evaluation of Random Forest model\n\n")
y_pred_rf = rf_clf.predict(X_test)
from sklearn.metrics import classification_report, confusion_matrix, accuracy_score
result = confusion_matrix(y_test, y_pred_rf)
print(f"Confusion Matrix:\n{result}")
result1 = classification_report(y_test, y_pred_rf)
print(f"Classification Report:\n{result1}")
result2 = accuracy_score(y_test,y_pred_rf)
print(f"Accuracy:{round(result2,2)}\n\n")
Evaluation of Random Forest model
Confusion Matrix:
[[ 790 66 31 61 9 604]
[ 33 1522 0 6 1 29]
[ 19 1 1546 1 0 12]
[ 124 6 4 1293 5 116]
[ 8 7 0 2 1532 18]
[ 442 10 19 84 11 964]]
Classification Report:
precision recall f1-score support
0 0.56 0.51 0.53 1561
1 0.94 0.96 0.95 1591
2 0.97 0.98 0.97 1579
3 0.89 0.84 0.86 1548
4 0.98 0.98 0.98 1567
5 0.55 0.63 0.59 1530
accuracy 0.82 9376
macro avg 0.82 0.81 0.81 9376
weighted avg 0.82 0.82 0.82 9376
Accuracy:0.82
因此，具有最佳精度、召回率和 f1 分数以及 82% 整体准确度的 最佳模型 是 随机森林分类器。