Text Sentiment Classification Experiment with Emoji

some thoughts

This is my content security class assignment. I refer to some online codes and articles. The top leaders in the report are all kinds of in-depth learning curve analysis and parameter adjustment. , can only choose a simple paper to reproduce, and make some improvements, which are recorded here to avoid stepping on the pit.

Knowledge points involved

Chinese word segmentation (part-of-speech analysis), TF-IDF, Naive Bayesian, artificial neural network

Thesis content

Entering the text, first reproduce the thesis, this paper is based on the emoticon analysis method of emotional key sentence extraction, which uses the sentence emotional polarity calculation, keyword calculation, and position information calculation based on emoji analysis
. Sentence emotion polarity calculation is an improvement to the existing method. By incorporating emoji analysis, the accuracy of emotion key sentence extraction is improved. The paper is as follows. This article mainly

introduces several new methods and new models of emoji classification, and does not propose implementation The specific way, so some changes were made when the code was implemented.

paper reproduction

Emoji processing

The first is to extract the emoticons, because the emoticons are all wrapped in [], so the matching and extraction are performed according to this feature, and all the emoticons in the data set are stored in an array for later processing. The first is the first function Introduction to pre_processer.py

Then store these emoticons to avoid repeated reading later and waste time

The next function classifies expressions. Here, refer to the co-occurrence rate judgment method in the paper to judge the emotion of unknown expressions. However, it should be noted that this is different from the recurrence basis in the paper. The data set in the paper uses word vectors. The training set, and the label information of the sentence vector is given here, so the method has been adjusted to judge the number of times the expression appears in positive sentences and negative sentences, and there are also neutral sentences

When the positive ratio is greater than 0.8, it is a positive expression, and the same is true for negative expressions. The rest are classified as neutral expressions because the emotions are not obvious, and are stored in three arrays. So far, the emotion classification of expressions has been completed.

text processing

The following is the part of word processing, that is, the specialize.py part, which reproduces some operations on text weights in the paper and makes some improvements. First of all, since the given data set is a sentence, and there are many impurities such as @, //, which are not good for analysis, the first thing to do

is to clean the data. First, perform matching filtering and regular filtering, and put some frequencies that are visible to the naked eye. Remove the high meaningless symbols

and then use the stop word filter

to find some deficiencies in the previous stop word list. The specific stop word list can be found on Baidu, because there are some expressions such as "very" and "special" that represent feelings Words that are filtered out here will affect the subsequent analysis, so all emotional words are manually deleted, leaving only a part of filtered useless words and symbols, followed by jieba word segmentation, which is used as part-of-speech tagging followed by the routine program TF-

IDF Select the important word components of each sentence. Here, select 5 keywords for each sentence to avoid too slow training later.

This also uses the weight adjustment of the importance of adverbs and the importance of positions mentioned in the paper, such as Say the following weight array

and modify the weight, for example, write a simple weight return function

In addition, the weight of the tail words is doubled compared with that of the middle words, and a speciallist feature vector is generated, which will be added during training later. This part is the training of words.

Build model training

Then there is the part of training. The first is to read the data set, define the initial weight of

positive and negative and neutral words and expression weights as 1, 0, and -1 respectively, and then directly perform Naive Bayesian training, which is convenient for later To compare the experimental results, it is worth mentioning that in the article, the words are scored separately to calculate the weight, but the data set does not meet the conditions when it is reproduced here, so the basic weight is given, that is, the weight of the sentence where the word is located exceeds 80%.

The 60,000 data sets here are cross-selected data sets and training sets, and the database is adjusted for scoring. The

classification is a weight correction of ordinary naive Bayesian.

It can be seen that the feature vectors processed before are added, and they are correct. Comparing the results and directly calculating the ratio, it can be seen from the results that the improvement is very obvious, from more than 50% to more than 80%, indicating that the method proposed in the paper is obviously effective (the results are written later).

The above are all definition methods, and the main function is as follows.

The three types are trained and superimposed separately. Because the effect of using artificial neural network in Experiment 2 is not bad, so here is also added an artificial neural network for processing.

I didn't use tools to adjust the parameters. I wrote two loops and ran it all night. I wrote the file to judge the parameter optimization by myself, and got the parameter combination with a higher score. The result was also significantly optimized.

result

Run the code and get the running result:

It can be seen that the effect of sentiment analysis on text data classification with expression assistance is better than that of naive Bayesian classification alone.

And deep learning is better than naive Bayesian

the code

Expression processing

import csv
# 获取某一评论中存在的表情 参数：path：文件路径 emotion_set：用来存储表情列表
def get_emotion(path, emotion_set):
    with open(path, encoding= 'utf-8-sig') as f:
        reader = csv.reader(f)
        rows=[row for row in  reader]
        for each in rows:
            for i in range(len(each[0])):
                if each[0][i] == '[' : # 判断是否为表情符号
                    temp = ''
                    for k in range(100):
                        if i+k > len(each[0]) - 1:
                            break
                        temp = temp + each[0][i+k]
                        # print(each[i+k])
                        if each[0][i+k] == ']':
                            if temp not in emotion_set:
                                emotion_set.append(temp) # 若为表情符号则存储在该列表中，使每个符号只出现一次
                            # print(temp)
                            break
    return emotion_set


# 整合三个文件获取的表情存储于emotion_result_sat，存储于csv
def Save_as_File(save_path, emotion_result_set):  
    emotion_set = []
    Star_emotion_set = get_emotion('明星.csv', emotion_set)
    Hotspot_emotion_set = get_emotion('热点.csv', emotion_set)
    Epidemic_emotion_set = get_emotion('疫情.csv', emotion_set)
    for each in Star_emotion_set:
        if each not in emotion_result_set:
            emotion_result_set.append(each)
    for each in Hotspot_emotion_set:
        if each not in emotion_result_set:
            emotion_result_set.append(each)
    for each in Epidemic_emotion_set:
        if each not in emotion_result_set:
            emotion_result_set.append(each)
            
    with open(save_path, 'w', encoding='utf-8-sig') as f:
        for each in emotion_result_set:
            write=csv.writer(f)
            write.writerow(each)
    return emotion_result_set


# 根据评论标签1或-1的多少判断表情是积极还是负面，积极的存储于PE_set,负面的存储于NE_set
def creat_Respiratory(emotion_result_set):
    PE_set = []
    NE_set = []
    Neu_set = []
    with open('明星.csv',encoding= 'utf-8-sig') as f1:
        reader1 = csv.reader(f1)
        rows1=[row for row in  reader1]      
        with open('热点.csv',encoding= 'utf-8-sig') as f2:
            reader2 = csv.reader(f2)
            rows2=[row for row in  reader2]  
            with open('疫情.csv',encoding= 'utf-8-sig') as f3:
                reader3 = csv.reader(f3)
                rows3=[row for row in  reader3]
        
                temp_set = []
                for a in rows1:
                    temp_set.append(a)
                for b in rows2:
                    temp_set.append(b)
                for c in rows3:
                    temp_set.append(c) #将所有评论存储在一个列表temp_set内
    # print(temp_set)
    for emotion in emotion_result_set:
        positive = 0
        negtive = 0
        for critic in temp_set:
            # print(emotion)
            if emotion in critic[0]:
                if critic[1] == '1':
                    positive = positive + 1
                if critic[1] == '-1':
                    negtive = negtive + 1
        if positive + negtive == 0:
            Neu_set.append(emotion)
        else:
            if (positive/(positive + negtive)) > 0.8:
                PE_set.append(emotion)
            elif (negtive/(positive + negtive)) > 0.8:
                NE_set.append(emotion)
            else:
                Neu_set.append(emotion)
    with open('Positive_emotion.csv', 'w', encoding='utf-8-sig') as f:
        for each in PE_set:
            write=csv.writer(f)
            write.writerow(each)
        
    with open('Negtive_emotion.csv', 'w', encoding='utf-8-sig') as f:
        for each in NE_set:
            write=csv.writer(f)
            write.writerow(each)
        
    with open('Neu_emotion.csv', 'w', encoding='utf-8-sig') as f:
        for each in Neu_set:
            write=csv.writer(f)
            write.writerow(each)

    return PE_set, NE_set, Neu_set
    
    

# emotion_result_set = []
# emotion_result_set = Save_as_File('.//emotion//All_emotion.csv', emotion_result_set)
# PE_set, NE_set = creat_Respiratory(emotion_result_set)

text processing

import pre_professer
import jieba
import csv
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfTransformer

list3=['最','最为','极','极其','极为','极度']
list2_5=['太','至','至为','顶','过','过于','过份','分外','万分']
list2=['很','挺','怪','非常','特别','相当','十分','甚为','够','多','多么']
list1_5=['不甚','不胜','好','好不','颇','颇为','大','大为']
list1_1=['稍','比较','较为','还']
list0_8=['稍稍','稍微','稍许','略微','多少']
list0_5=['有点','有些']
list_1=['甭', '别', '不', '不曾', '不必', '非', '没', '没有', '莫', '未必', '未尝', '无从', '无须', '不要', '不用', '不再', '不很', '不太', '绝非', '决非', '并非', '不能', '不常', '不会', '不可能', '何曾', '何尝', '勿']

def PLfenci(emotion_result_set, _class): #遍历每一条评论分词处理并去杂(去除表情、去除@内容、去除停用词)
    label_list = [] # 一维列表，存储评论标签
    with open('明星.csv',encoding= 'utf-8-sig') as f1:
        reader1 = csv.reader(f1)
        rows1=[row for row in  reader1]      
        with open('热点.csv',encoding= 'utf-8-sig') as f2:
            reader2 = csv.reader(f2)
            rows2=[row for row in  reader2]  
            with open('疫情.csv',encoding= 'utf-8-sig') as f3:
                reader3 = csv.reader(f3)
                rows3=[row for row in  reader3]
        
                temp_set = []
                for a in rows1:
                    if _class == 'Star' or _class == 'All':
                        temp_set.append(a)
                    label_list.append(a[1])
                for b in rows2:
                    if _class == 'Hotspot' or _class == 'All':
                        temp_set.append(b)
                    label_list.append(b[1])
                for c in rows3:
                    if _class == 'Epidemic' or _class == 'All':
                        temp_set.append(c) #将评论存储在一个列表temp_set内
                    label_list.append(c[1])
                if _class == 'All':
                    return temp_set, label_list
# 去除表情，去除@信息和停用表词
    for i in range(len(temp_set)):
        for emotion in emotion_result_set:
            if emotion in temp_set[i][0]:
                temp_set[i][0] = temp_set[i][0].replace(emotion, '')

    for each in temp_set:
        for m in range(len(each[0])):
            if (each[0][m] == '/') and (each[0][m+1] == '/') and (each[0][m+2] == '@'):
                while m < len(each[0]):
                    each[0] = each[0].replace(each[0][m], '')
                    m+=1
                break    
            
    for each in temp_set:
        for i in range(len(each[0])):
            if i < len(each[0]) and each[0][i] == '@':
                temp = ''
                while i < len(each[0]) and (each[0][i] != '：' or each[0][i] == ' '): # 注意要使用中文'：'
                    temp = temp + each[0][i]
                    i = i + 1
                each[0] = each[0].replace(temp, '')
           
    stop_list = []
    f_stop = open('.\\停用词表.txt', 'r', encoding = 'UTF-8')
    #获取停用词列表
    for each in f_stop:
        each = each.strip()#去除尾字符中的换行符
        stop_list.append(each)
    for each in temp_set:
        for k in stop_list:
            if k in each[0]:
                each[0] = each[0].replace(k, '')
    
# 进行jieba分词
    speciallist=[]
    data_list = [] # 二维列表，每个元素为一个评论预处理后的结果，每个结果也是一个列表，用','分割。
    for k in range(len(temp_set)):
        data_list.append([])
    k = 0
    for each in temp_set:
        a=checkpoint(each)
        speciallist.append(a)
        generator = jieba.cut(each[0])
        for i in generator:
            data_list[k].append(i)
        k = k + 1
    

# 预处理结果写入segment文件
    with open('.\\segment.csv', 'w', newline = '',encoding = 'utf-8-sig') as f:
        write = csv.writer(f)
        for each in data_list:
            write.writerow(each)

    return data_list, label_list,speciallist


def tfidf_get(data_list, _class): # 实现3，4功能 data_list即为前面获取的数据列表, 类名（和前一函数一致）
    transfer_data_list = [] #将data_list转化为tfidf可以看懂的格式
    for each in data_list:
        x = ' '.join(each)
        transfer_data_list.append(x)
    vectorizer = CountVectorizer() # 将文本中的词语转换为词频矩阵
    X = vectorizer.fit_transform(transfer_data_list) # 计算个词语出现的次数
    word_list = vectorizer.get_feature_names()  # 获取词袋中所有文本关键词
    transformer = TfidfTransformer() #类调用
    tfidf = transformer.fit_transform(X) #将词频矩阵X统计成TF-IDF值
    feature = [] # 存储所有评论的feature， 每个元素为一条评论的feature，每条评论5个feature。
    length = len(word_list)
    s = 0
    for each in tfidf.toarray():#进入二维列表tfidf_list的每一项中，遍历每一句的tfidf
        feature_list = []    #对于每一评论，初始特征置空
        item = 0#特征值计数变量，满5个为止
        while item < 5:
            max_tfidf = 0#最大值置零
            for i in range(length):#每一次循环找出一个最大tfidf值对应的word并存入feature_list列表中，并置该tfidf = 0
                if each[i] >= max_tfidf:
                    max_tfidf = each[i] 
                    m = i#记录相应索引号
                if i == (length - 1):# i = length - 1说明句子已经遍历到末尾了，此时最大的tfidf值就可以确定了
                    feature_list.append(word_list[m])
                    each[m] = 0
                    item = item + 1
        v = ' '.join(feature_list)
        feature.append(v)
        s = s + 1
    with open('.\\特征数据\\' + _class + '.csv', 'w', newline = '',encoding = 'utf-8-sig') as f:
        write = csv.writer(f)
        for each in feature:
            write.writerow(each)

def checkpoint(str):
    if str in list3:
        return 3
    elif str in list2_5:
        return 2.5
    elif str in list2:
        return 2
    elif str in list1_5:
        return 1.5
    elif str in list1_1:
        return 1.1
    elif str in list0_8:
        return 0.8
    elif str in list0_5:
        return 0.5
    elif str in list_1:
        return -1
    else:
        return 0

modeling training

import pre_professer
import specialize
from sklearn.model_selection import  train_test_split #数据集划分
from sklearn.naive_bayes import MultinomialNB
from sklearn.feature_extraction.text import CountVectorizer  # 从sklearn.feature_extraction.text里导入文本特征向量化模块
from sklearn.metrics import classification_report
from sklearn.neural_network import MLPClassifier
import time

data = []#数据集
def Contribute_data(data, _class): # 获取class类别的数据
    with open('\\特征数据\\' + _class + '.csv', 'r', encoding = 'utf-8-sig') as fr:
        for tem in fr:
            tem = tem.replace(',', '')
            tem = tem.replace('\n', '')
            data.append(tem)
    return ''  
def get_target(lebal_list):
    target = []
    for each in lebal_list:
        if each == '1':
            target.append('积极')
        elif each == '0':
            target.append('中性')
        elif each == '-1':
            target.append('消极')
        else:
            target.append('中性')
    return target

def Bayesian(data, target):
    #数据预处理：训练集和测试集分割，文本特征向量化
    #X_train,X_test,y_train,y_test = train_test_split(data, target, test_size=30000 ,random_state=4) # 随机采样数据样本作为测试集
    X_train = data[10000:20000] + data[30000:40000] + data[60000: 80000]
    X_test = data[0:10000] + data[20000:30000] + data[50000: 60000]
    y_train = target[10000:20000] + target[30000:40000] + target[60000: 80000]
    y_test = target[0:10000] + target[20000:30000] + target[50000: 60000]
    k = X_test
    #文本特征向量化
    vec = CountVectorizer()
    X_train = vec.fit_transform(X_train)
    X_test = vec.transform(X_test)
    #print(y_test)
    #print(y_train)
    #使用朴素贝叶斯进行训练
    mnb = MultinomialNB()   # 使用默认配置初始化朴素贝叶斯
    mnb.fit(X_train,y_train)    # 利用训练数据对模型参数进行估计
    y_predict = mnb.predict(X_test)     # 对参数进行预测
    #print(y_predict)
    #获取结果报告
    print ('不带表情的朴素贝叶斯:', mnb.score(X_test,y_test))
    print ('其它指标:\n',classification_report(y_test, y_predict, target_names = ['积极', '中性', '消极']))
    
    return k, y_test, y_predict

def Fenlei(emotion_X_test, emotion_y_test, emotion_y_predict, origin_list, PE_set, NE_set, Neu_set,specialist): # 运用表情训练
    o_list = origin_list[0:10000] + origin_list[20000:30000] + origin_list[50000: 60000]
    length = len(emotion_y_predict)
    for i in range(length):
        if emotion_y_test[i] != emotion_y_predict[i]:
            key = 0
            if '' in emotion_X_test[i]:

                for each in PE_set:
                    if each in o_list[i]:
                        key = key + 1
                for each in NE_set:
                    if each in o_list[i]:
                        key = key - 1
                key+=specialist[i]
                if key > 0:
                    emotion_y_predict[i] = '积极'
                if key < 0:
                    emotion_y_predict[i] = '消极'
                if key == 0:
                   emotion_y_predict[i] = '中性'
            else:
                key = 0.5
                for each in PE_set:
                    if each in o_list[i]:
                        key = key + 1
                for each in NE_set:
                    if each in o_list[i]:
                        key = key - 1
                for each in Neu_set:
                    if each in o_list[i]:
                        key = key - 0.5
                if key > 0:
                    emotion_y_predict[i] = '积极'
                if key < 0:
                    emotion_y_predict[i] = '消极'
                if key == 0:
                    emotion_y_predict[i] = '中性'
# 呈现分类结果
    rate = 0
    for p in range(length):
        if emotion_y_predict[p] == emotion_y_test[p]:
            rate += 1
    print('带表情的贝叶斯：%.8f'%(rate/30000))
                
    return ''
start=time.time()
print('开始预处理')
emotion_result_set = []
emotion_result_set = pre_professer.Save_as_File('All_emotion.csv', emotion_result_set)
PE_set, NE_set, Neu_set = pre_professer.creat_Respiratory(emotion_result_set)
print('预处理完毕')

Star_data_list,lebal_list ,specialkist= specialize.PLfenci(emotion_result_set, 'Star')
print('开始star处理')
feature = specialize.tfidf_get(Star_data_list, 'Star')
print('Star类别特征构建成功')
print('开始hotpot处理')
Hotspot_data_list,lebal_list ,specialkist= specialize.PLfenci(emotion_result_set, 'Hotspot')
feature = specialize.tfidf_get(Hotspot_data_list, 'Hotspot')
print('Hotspot特征构建成功')
Epidemic_data_list,lebal_list ,specialkist= specialize.PLfenci(emotion_result_set, 'Epidemic')
feature = specialize.tfidf_get(Epidemic_data_list, 'Epidemic')
print('Epidemic特征构建成功')

Contribute_data(data, 'Star')
Contribute_data(data, 'Hotspot')
Contribute_data(data, 'Epidemic')

origin_list, lebal_list = specialize.PLfenci(emotion_result_set, 'All')
target = get_target(lebal_list)
emotion_X_test, emotion_y_test, emotion_y_predict = Bayesian(data, target)
#print(len(emotion_y_test))
Fenlei(emotion_X_test, emotion_y_test, emotion_y_predict, origin_list, PE_set, NE_set, Neu_set,specialkist)
size = (10,10000)
iters = 10000
clf = MLPClassifier(activation='relu', alpha=1e-05, batch_size='auto', beta_1=0.9,
       beta_2=0.999, early_stopping=False, epsilon=1e-08,
       hidden_layer_sizes=size, learning_rate='constant',
       learning_rate_init=0.001, max_iter=iters, momentum=0.9,
       nesterovs_momentum=True, power_t=0.5, random_state=1, shuffle=True,
       solver='adam', tol=0.0001, validation_fraction=0.1, verbose=False,
       warm_start=False)#多层感知机
tv=specialize.tfidf_get()
print('深度学习:',clf.score(tv.transform(emotion_X_test),emotion_y_test))
from sklearn.metrics import classification_report
print('其它指标:\n',classification_report(emotion_y_predict,clf.predict(tv.transform(emotion_X_test))))
end=time.time()

Reference paper: Emotional key sentence extraction method based on emoticon analysis
(extraction code: 3882)