Python利用卡方Chi特征检验实现提取关键文本特征
Toblerone_Wind 人气:0理论
类别classi | 非类别classi | |
包含单词wordj的文档数 | A | B |
不包含单词wordj的文档数 | C | D |
卡方特征提取主要度量类别classi和单词wordj之间的依赖关系。计算公式如下
其中N是文档总数,A是包含单词wordj且属于classi的文档数,B是包含单词wordj但不属classi的文档数,C是不包含单词wordj但属于classi的文档数,D是不包含单词wordj且不属于classi的文档数。值得注意的是
最终单词wordj的CHI值计算公式如下,其中P(classi)表示属于类别 classi的文档在所有文档中出现的概率,k为总的类别数
代码
下面以二分类为例介绍一段python代码:第一个参数是文档列表,包含若干个文档,每个文档由若干个单词通过空格拼接而成;第二个参数是标签列表,对应每个文档的类别;第三个参数用来确定选取前百分之多少的单词。
# documents = [document_1, document_2, document_3, ...] # document_i = "word_1 word_2 word_3" # labels is a list combined with 0 and 1 def feature_word_select(documents:list, labels:list, percentage:float): # get all words word_set = set() for document in documents: words = document.split() word_set.update(words) word_list = list(word_set) word_list.sort() sorted_words = chi(word_list, documents, labels) top_k_words = sorted_words[:int(percentage * len(sorted_words))] return top_k_words
这段代码首先创建一个集合word_set,接着遍历所有的文档,对每一个文档,使用split()函数对其进行切分,得到一个words列表,再将列表中的所有元素输入到集合word_set中,word_set由于集合的特性会过滤集合中已有的单词。收集完毕后,通过word_set生成一个单词列表word_list。
将单词列表,文档列表和标签列表输入chi函数,得到通过卡方值降序排列的单词列表sorted_words。
最后选取前百分之percentage的单词最为最后的特征单词。
下面这个函数cal_chi_word_class()用来计算 CHI(word, 0)和CHI(word, 1)。这里的A1表示属于类别1的A,A0表示属于类别0的A。
值得说明的是,在二分类问题中,A1实际上等于B0,C1实际上等于D0。因此,仅计算A1,B1,C1,D1即可推导出A0,B0,C0,D0。
此外,由于文档总数N对于CHI(word, 0)和CHI(word, 1)来说属于公共的分子且保持不变,所以可以不参与计算;A1+C1=B0+D0,B1+D1=A0+C0,所以CHI(word, 0)和CHI(word, 1)的分母部分可以进行简化
# calculate chi(word,1) and chi(word,0) def cal_chi_word_class(word, labels, documents): N = len(documents) A1, B1, C1, D1 = 0., 0., 0., 0. A0, B0, C0, D0 = 0., 0., 0., 0. for i in range(len(documents)): if word in documents[i].split(): if labels[i] == 1: A1 += 1 B0 += 1 else: B1 += 1 A0 += 1 else: if labels[i] == 1: C1 += 1 D0 += 1 else: D1 += 1 C0 += 1 chi_word_1 = N * (A1*D1-C1*B1)**2 / ((A1+C1)*(B1+D1)*(A1+B1)*(C1+D1)) chi_word_0 = N * (A0*D0-C0*B0)**2 / ((A0+C0)*(B0+D0)*(A0+B0)*(C0+D0)) return chi_word_1, chi_word_0
简化后
# calculate chi(word,1) and chi(word,0) def cal_chi_word_class(word, labels, documents): A1, B1, C1, D1 = 0., 0., 0., 0. for i in range(len(documents)): if word in documents[i].split(): if labels[i] == 1: A1 += 1 else: B1 += 1 else: if labels[i] == 1: C1 += 1 else: D1 += 1 A0, B0, C0, D0 = B1, A1, D1, C1 chi_word_1 = (A1*D1-C1*B1)**2 / ((A1+B1)*(C1+D1)) chi_word_0 = (A0*D0-C0*B0)**2 / ((A0+B0)*(C0+D0)) return chi_word_1, chi_word_0
在chi函数中调用cal_chi_word_class函数,即可计算每个单词的卡方值,以字典的形式保存每个单词的卡方值,最后对字典的所有值进行排序,并提取出排序后的单词。
def chi(word_list, documents, labels): P1 = labels.count(1) / len(documents) P0 = 1 - P1 dic = {} for word in word_list: chi_word_1, chi_word_0 = cal_chi_word_class(word, labels, documents) chi_word = P0 * chi_word_0 + P1 * chi_word_1 dic[word] = chi_word sorted_list = sorted(dic.items(), key=lambda x:x[1], reverse=True) sorted_chi_word = [x[0] for x in sorted_list] return sorted_chi_word
测试代码。这里我略过了数据处理环节,documents列表中的每一个元素document_i都是有若干个单词或符号通过空格拼接而成。
def main(): documents = ["today i am happy !", "she is not happy at all", "let us go shopping !", "mike was so sad last night", "amy did not love it", "it is so amazing !" ] labels = [1, 0, 1, 0, 0, 1] words = feature_word_select(documents, labels, 0.3) print(words) if __name__ == '__main__': main()
运行结果如下
['!', 'not', 'all', 'am', 'amazing', 'amy', 'at']
进一步,可以在chi函数里输出sorted_list(每个单词对应的卡方值),结果如下。这里输出的并不是真实的卡方值,是经过化简的,如需输出原始值,请使用完整版的cal_chi_word_class()函数。
[('!', 9.0), ('not', 4.5), ('all', 1.8), ('am', 1.8), ('amazing', 1.8), ('amy', 1.8), ('at', 1.8), ('did', 1.8), ('go', 1.8), ('i', 1.8), ('last', 1.8), ('let', 1.8), ('love', 1.8), ('mike', 1.8), ...]
完整代码
# calculate chi(word,1) and chi(word,0) def cal_chi_word_class(word, labels, documents): A1, B1, C1, D1 = 0., 0., 0., 0. for i in range(len(documents)): if word in documents[i].split(): if labels[i] == 1: A1 += 1 else: B1 += 1 else: if labels[i] == 1: C1 += 1 else: D1 += 1 A0, B0, C0, D0 = B1, A1, D1, C1 chi_word_1 = (A1*D1-C1*B1)**2 / ((A1+B1)*(C1+D1)) chi_word_0 = (A0*D0-C0*B0)**2 / ((A0+B0)*(C0+D0)) return chi_word_1, chi_word_0 def chi(word_list, documents, labels): P1 = labels.count(1) / len(documents) P0 = 1 - P1 dic = {} for word in word_list: chi_word_1, chi_word_0 = cal_chi_word_class(word, labels, documents) chi_word = P0 * chi_word_0 + P1 * chi_word_1 dic[word] = chi_word sorted_list = sorted(dic.items(), key=lambda x:x[1], reverse=True) sorted_chi_word = [x[0] for x in sorted_list] return sorted_chi_word # documents = [document_1, document_2, document_3, ...] # document_i = "word_1 word_2 word_3" # labels is a list combined with 0 and 1 def feature_word_select(documents:list, labels:list, percentage:float): # get all words word_set = set() for document in documents: words = document.split() word_set.update(words) word_list = list(word_set) word_list.sort() sorted_words = chi(word_list, documents, labels) top_k_words = sorted_words[:int(percentage * len(sorted_words))] return top_k_words def main(): documents = ["today i am happy !", "she is not happy at all", "let us go shopping !", "mike was so sad last night", "amy did not love it", "it is so amazing !" ] labels = [1, 0, 1, 0, 0, 1] words = feature_word_select(documents, labels, 0.3) print(words) if __name__ == '__main__': main()
题外话
卡方特征选择仅考虑单词是否在文档中出现,并没有考虑单词出现的次数,因此选择出的特征单词可能并不准确。
加载全部内容