原理我就不讲了,请移步下面这篇论文,包括情感词典的构建(各位读者可以根据自己的需求稍作简化),以及打分策略(程序对原论文稍有改动)。
论文在这里下载:基于情感词典的中文微博情感倾向性研究
(大家可以上知网自行下载)
本文采用的方法如下:
首先对单条微博进行文本预处理,并以标点符号为分割标志,将单条微博分割为n个句子,提取每个句子中的情感词 。以下两步的处理均以分句为处理单位。
第二步在情感词表中寻找情感词,以每个情感词为基准,向前依次寻找程度副词、否定词,并作相应分值计算。随后对分句中每个情感词的得分作求和运算。
第三步判断该句是否为感叹句,是否为反问句,以及是否存在表情符号。如果是,则分句在原有分值的基础上加上或减去对应的权值。
最后对该条微博的所有分句的分值进行累加,获得该条微博的最终得分。
代码如下:
首先文件结构图如下:
其中,degree_dict为程度词典,其中每个文件为不同的权值。
emotion_dict为情感词典,包括了积极情感词和消极情感词以及停用词。
文件一:文本预处理 textprocess.py
在里面封装了一些文本预处理的函数,方便调用。
# -*- coding: utf-8 -*-
__author__ = 'Bai Chenjia'
import jieba
import jieba.posseg as pseg
print "加载用户词典..."
import sys
reload(sys)
sys.setdefaultencoding("utf-8")
jieba.load_userdict('C://Python27/Lib/site-packages/jieba/user_dict/pos_dict.txt')
jieba.load_userdict('C://Python27/Lib/site-packages/jieba/user_dict/neg_dict.txt')
# 分词,返回List
def segmentation(sentence):
seg_list = jieba.cut(sentence)
seg_result = []
for w in seg_list:
seg_result.append(w)
#print seg_result[:]
return seg_result
# 分词,词性标注,词和词性构成一个元组
def postagger(sentence):
pos_data = pseg.cut(sentence)
pos_list = []
for w in pos_data:
pos_list.append((w.word, w.flag))
#print pos_list[:]
return pos_list
# 句子切分
def cut_sentence(words):
words = words.decode('utf8')
start = 0
i = 0
token = 'meaningless'
sents = []
punt_list = ',.!?;~,。!?;~… '.decode('utf8')
#print "punc_list", punt_list
for word in words:
#print "word", word
if word not in punt_list: # 如果不是标点符号
#print "word1", word
i += 1
token = list(words[start:i+2]).pop()
#print "token:", token
elif word in punt_list and token in punt_list: # 处理省略号
#print "word2", word
i += 1
token = list(words[start:i+2]).pop()
#print "token:", token
else:
#print "word3", word
sents.append(words[start:i+1]) # 断句
start = i + 1
i += 1
if start < len(words): # 处理最后的部分
sents.append(words[start:])
return sents
def read_lines(filename):
fp = open(filename, 'r')
lines = []
for line in fp.readlines():
line = line.strip()
line = line.decode("utf-8")
lines.append(line)
fp.close()
return lines
# 去除停用词
def del_stopwords(seg_sent):
stopwords = read_lines("f://Sentiment_dict/emotion_dict/stop_words.txt") # 读取停用词表
new_sent = [] # 去除停用词后的句子
for word in seg_sent:
if word in stopwords:
continue
else:
new_sent.append(word)
return new_sent
# 获取六种权值的词,根据要求返回list,这个函数是为了配合Django的views下的函数使用
def read_quanzhi(request):
result_dict = []
if request == "one":
result_dict = read_lines("f://emotion/mysite/Sentiment_dict/degree_dict/most.txt")
elif request == "two":
result_dict = read_lines("f://emotion/mysite/Sentiment_dict/degree_dict/very.txt")
elif request == "three":
result_dict = read_lines("f://emotion/mysite/Sentiment_dict/degree_dict/more.txt")
elif request == "four":
result_dict = read_lines("f://emotion/mysite/Sentiment_dict/degree_dict/ish.txt")
elif request == "five":
result_dict = read_lines("f://emotion/mysite/Sentiment_dict/degree_dict/insufficiently.txt")
elif request == "six":
result_dict = read_lines("f://emotion/mysite/Sentiment_dict/degree_dict/inverse.txt")
else:
pass
return result_dict
if __name__ == '__main__':
test_sentence1 = "这款手机大小合适。"
test_sentence2 = "这款手机大小合适,配置也还可以,很好用,只是屏幕有点小。。。总之,戴妃+是一款值得购买的智能手机。"
test_sentence3 = "这手机的画面挺好,操作也比较流畅。不过拍照真的太烂了!系统也不好。"
"""
seg_result = segmentation(test_sentence3) # 分词,输入一个句子,返回一个list
for w in seg_result:
print w,
print '
'
"""
"""
new_seg_result = del_stopwords(seg_result) # 去除停用词
for w in new_seg_result:
print w,
"""
#postagger(test_sentence1) # 分词,词性标注,词和词性构成一个元组
#cut_sentence(test_sentence2) # 句子切分
#lines = read_lines("f://Sentiment_dict/emotion_dict/posdict.txt")
#print lines[:]
文件二:情感打分 dict_main.py
其中待处理数据放在chinese_weibo.txt中,读者可以自行更改文件目录,该文件中的数据格式如下图:
即用每一行代表一条语句,我们对每条语句进行情感分析,进行打分
# -*- coding: utf-8 -*-
__author__ = 'Bai Chenjia'
import text_process as tp
import numpy as np
# 1.读取情感词典和待处理文件
# 情感词典
print "reading..."
posdict = tp.read_lines("f://emotion/mysite/Sentiment_dict/emotion_dict/pos_all_dict.txt")
negdict = tp.read_lines("f://emotion/mysite/Sentiment_dict/emotion_dict/neg_all_dict.txt")
# 程度副词词典
mostdict = tp.read_lines('f://emotion/mysite/Sentiment_dict/degree_dict/most.txt') # 权值为2
verydict = tp.read_lines('f://emotion/mysite/Sentiment_dict/degree_dict/very.txt') # 权值为1.5
moredict = tp.read_lines('f://emotion/mysite/Sentiment_dict/degree_dict/more.txt') # 权值为1.25
ishdict = tp.read_lines('f://emotion/mysite/Sentiment_dict/degree_dict/ish.txt') # 权值为0.5
insufficientdict = tp.read_lines('f://emotion/mysite/Sentiment_dict/degree_dict/insufficiently.txt') # 权值为0.25
inversedict = tp.read_lines('f://emotion/mysite/Sentiment_dict/degree_dict/inverse.txt') # 权值为-1
# 情感级别
emotion_level1 = "悲伤。在这个级别的人过的是八辈子都懊丧和消沉的生活。这种生活充满了对过去的懊悔、自责和悲恸。在悲伤中的人,看这个世界都是灰黑色的。"
emotion_level2 = "愤怒。如果有人能跳出冷漠和内疚的怪圈,并摆脱恐惧的控制,他就开始有欲望了,而欲望则带来挫折感,接着引发愤怒。愤怒常常表现为怨恨和复仇心里,它是易变且危险的。愤怒来自未能满足的欲望,来自比之更低的能量级。挫败感来自于放大了欲望的重要性。愤怒很容易就导致憎恨,这会逐渐侵蚀一个人的心灵。"
emotion_level3 = "淡定。到达这个能级的能量都变得很活跃了。淡定的能级则是灵活和无分别性的看待现实中的问题。到来这个能级,意味着对结果的超然,一个人不会再经验挫败和恐惧。这是一个有安全感的能级。到来这个能级的人们,都是很容易与之相处的,而且让人感到温馨可靠,这样的人总是镇定从容。他们不会去强迫别人做什么。"
emotion_level4 = "平和。他感觉到所有的一切都生机勃勃并光芒四射,虽然在其他人眼里这个世界还是老样子,但是在这人眼里世界却是一个。所以头脑保持长久的沉默,不再分析判断。观察者和被观察者成为同一个人,观照者消融在观照中,成为观照本身。"
emotion_level5 = "喜悦。当爱变得越来越无限的时候,它开始发展成为内在的喜悦。这是在每一个当下,从内在而非外在升起的喜悦。这个能级的人的特点是,他们具有巨大的耐性,以及对一再显现的困境具有持久的乐观态度,以及慈悲。同时发生着。在他们开来是稀松平常的作为,却会被平常人当成是奇迹来看待。"
# 情感波动级别
emotion_level6 = "情感波动很小,个人情感是不易改变的、经得起考验的。能够理性的看待周围的人和事。"
emotion_level7 = "情感波动较大,周围的喜悦或者悲伤都能轻易的感染他,他对周围的事物有敏感的认知。"
# 2.程度副词处理,根据程度副词的种类不同乘以不同的权值
def match(word, sentiment_value):
if word in mostdict:
sentiment_value *= 2.0
elif word in verydict:
sentiment_value *= 1.75
elif word in moredict:
sentiment_value *= 1.5
elif word in ishdict:
sentiment_value *= 1.2
elif word in insufficientdict:
sentiment_value *= 0.5
elif word in inversedict:
#print "inversedict", word
sentiment_value *= -1
return sentiment_value
# 3.情感得分的最后处理,防止出现负数
# Example: [5, -2] → [7, 0]; [-4, 8] → [0, 12]
def transform_to_positive_num(poscount, negcount):
pos_count = 0
neg_count = 0
if poscount < 0 and negcount >= 0:
neg_count += negcount - poscount
pos_count = 0
elif negcount < 0 and poscount >= 0:
pos_count = poscount - negcount
neg_count = 0
elif poscount < 0 and negcount < 0:
neg_count = -poscount
pos_count = -negcount
else:
pos_count = poscount
neg_count = negcount
return (pos_count, neg_count)
# 求单条微博语句的情感倾向总得分
def single_review_sentiment_score(weibo_sent):
single_review_senti_score = []
cuted_review = tp.cut_sentence(weibo_sent) # 句子切分,单独对每个句子进行分析
for sent in cuted_review:
seg_sent = tp.segmentation(sent) # 分词
seg_sent = tp.del_stopwords(seg_sent)[:]
#for w in seg_sent:
# print w,
i = 0 # 记录扫描到的词的位置
s = 0 # 记录情感词的位置
poscount = 0 # 记录该分句中的积极情感得分
negcount = 0 # 记录该分句中的消极情感得分
for word in seg_sent: # 逐词分析
#print word
if word in posdict: # 如果是积极情感词
#print "posword:", word
poscount += 1 # 积极得分+1
for w in seg_sent[s:i]:
poscount = match(w, poscount)
#print "poscount:", poscount
s = i + 1 # 记录情感词的位置变化
elif word in negdict: # 如果是消极情感词
#print "negword:", word
negcount += 1
for w in seg_sent[s:i]:
negcount = match(w, negcount)
#print "negcount:", negcount
s = i + 1
# 如果是感叹号,表示已经到本句句尾
elif word == "!".decode("utf-8") or word == "!".decode('utf-8'):
for w2 in seg_sent[::-1]: # 倒序扫描感叹号前的情感词,发现后权值+2,然后退出循环
if w2 in posdict:
poscount += 2
break
elif w2 in negdict:
negcount += 2
break
i += 1
#print "poscount,negcount", poscount, negcount
single_review_senti_score.append(transform_to_positive_num(poscount, negcount)) # 对得分做最后处理
pos_result, neg_result = 0, 0 # 分别记录积极情感总得分和消极情感总得分
for res1, res2 in single_review_senti_score: # 每个分句循环累加
pos_result += res1
neg_result += res2
#print pos_result, neg_result
result = pos_result - neg_result # 该条微博情感的最终得分
result = round(result, 1)
return result
"""
# 测试
weibo_sent = "这手机的画面挺好,操作也比较流畅。不过拍照真的太烂了!系统也不好。"
score = single_review_sentiment_score(weibo_sent)
print score
"""
# 分析test_data.txt 中的所有微博,返回一个列表,列表中元素为(分值,微博)元组
def run_score():
fp_test = open('f://emotion/mysite/Weibo_crawler/chinese_weibo.txt', 'r') # 待处理数据
contents = []
for content in fp_test.readlines():
content = content.strip()
content = content.decode("utf-8")
contents.append(content)
fp_test.close()
results = []
for content in contents:
score = single_review_sentiment_score(content) # 对每条微博调用函数求得打分
results.append((score, content)) # 形成(分数,微博)元组
return results
# 将(分值,句子)元组按行写入结果文件test_result.txt中
def write_results(results):
fp_result = open('test_result.txt', 'w')
for result in results:
fp_result.write(str(result[0]))
fp_result.write(' ')
fp_result.write(result[1])
fp_result.write('
')
fp_result.close()
# 求取测试文件中的正负极性的微博比,正负极性分值的平均值比,正负分数分别的方差
def handel_result(results):
# 正极性微博数量,负极性微博数量,中性微博数量,正负极性比值
pos_number, neg_number, mid_number, number_ratio = 0, 0, 0, 0
# 正极性平均得分,负极性平均得分, 比值
pos_mean, neg_mean, mean_ratio = 0, 0, 0
# 正极性得分方差,负极性得分方差
pos_variance, neg_variance, var_ratio = 0, 0, 0
pos_list, neg_list, middle_list, total_list = [], [], [], []
for result in results:
total_list.append(result[0])
if result[0] > 0:
pos_list.append(result[0]) # 正极性分值列表
elif result[0] < 0:
neg_list.append(result[0]) # 负极性分值列表
else:
middle_list.append(result[0])
#################################各种极性微博数量统计
pos_number = len(pos_list)
neg_number = len(neg_list)
mid_number = len(middle_list)
total_number = pos_number + neg_number + mid_number
number_ratio = pos_number/neg_number
pos_number_ratio = round(float(pos_number)/float(total_number), 2)
neg_number_ratio = round(float(neg_number)/float(total_number), 2)
mid_number_ratio = round(float(mid_number)/float(total_number), 2)
text_pos_number = "积极微博条数为 " + str(pos_number) + " 条,占全部微博比例的 %" + str(pos_number_ratio*100)
text_neg_number = "消极微博条数为 " + str(neg_number) + " 条,占全部微博比例的 %" + str(neg_number_ratio*100)
text_mid_number = "中性情感微博条数为 " + str(mid_number) + " 条,占全部微博比例的 %" + str(mid_number_ratio*100)
##################################正负极性平均得分统计
pos_array = np.array(pos_list)
neg_array = np.array(neg_list) # 使用numpy导入,便于计算
total_array = np.array(total_list)
pos_mean = pos_array.mean()
neg_mean = neg_array.mean()
total_mean = total_array.mean() # 求单个列表的平均值
mean_ratio = pos_mean/neg_mean
if pos_mean <= 6: # 赋予不同的情感等级
text_pos_mean = emotion_level4
else:
text_pos_mean = emotion_level5
if neg_mean >= -6:
text_neg_mean = emotion_level2
else:
text_neg_mean = emotion_level1