1

我是 NLP 新手,请澄清如何使用 fit_transform 转换 TFIDF 值。

下面计算 IDF 的公式工作正常,log(文档总数 + 1 / 术语出现次数 + 1)+ 1

EG:文档 1 中术语“This”的 IDF 值(“this is a string”为 1.91629073

应用 fit_transform 后,所有项的值都发生了变化,用于转换的公式\逻辑是什么

TFID = TF * IDF

EG:文档 1(“this is a string”)中术语“This”的 TFIDF 值为 0.61366674

这个值是怎么来的,0.61366674?

from sklearn.feature_extraction.text import TfidfVectorizer
import pandas as pd

d = pd.Series(['This is a string','This is another string',
               'TFIDF Computation Calculation','TFIDF is the product of TF and IDF'])


df = pd.DataFrame(d)

tfidf_vectorizer = TfidfVectorizer()

tfidf = tfidf_vectorizer.fit_transform(df[0])


print (tfidf_vectorizer.idf_)

#output
#[1.91629073 1.91629073 1.91629073 1.91629073 1.91629073 1.22314355 1.91629073 
#1.91629073 1.51082562 1.91629073 1.51082562 1.91629073 1.51082562]

##-------------------------------------------------

##how the above values are getting transformed here 

##-------------------------------------------------


print (tfidf.toarray())


#[[0.         0.         0.         0.         0.         0.49681612  0.         
#0.         0.61366674 0.         0.         0.     0.61366674]
# [0.         0.61422608 0.         0.         0.         0.39205255
#  0.         0.         0.4842629  0.         0.         0.  0.4842629 ]
# [0.         0.         0.61761437 0.61761437 0.         0.
#  0.         0.         0.         0.         0.48693426 0.  0.        ]
# [0.37718389 0.         0.         0.         0.37718389 0.24075159
#  0.37718389 0.37718389 0.         0.37718389 0.29737611 0.37718389  0.        ]]
4

1 回答 1

0

它是规范的 TF-IDF 向量,因为默认情况下norm='l2'根据文档。因此,在tfidf.toarray()数组的第 0 层/行的每个元素的输出中,代表一个文档,第 1 层/列的每个元素代表一个唯一词,每个文档的向量元素的平方和等于 1,您可以检查通过印刷print([sum([word ** 2 for word in doc]) for doc in tfidf.toarray()])

norm : 'l1', 'l2' or None, optional (default='l2') 每个输出行都有单位范数,或者: * 'l2':向量元素的平方和为 1。两个向量之间的余弦相似度是应用 l2 范数时的点积。* 'l1':向量元素的绝对值之和为 1。见 preprocessing.normalize

print(tfidf) #the same values you find in tfidf.toarray() but more readable
output: ([index of document on array lvl 0 / row], [index of unique word on array lvl 1 / column]) normed TF-IDF value
(0, 12) 0.6136667440107333  #1st word in 1st sentence: 'This'
(0, 5)  0.4968161174826459  #'is'
(0, 8)  0.6136667440107333  #'string', see that word 'a' is missing
(1, 12) 0.48426290003607125 #'This'
(1, 5)  0.3920525532545391  #'is'
(1, 8)  0.48426290003607125 #'string'
(1, 1)  0.6142260844216119  #'another'
(2, 10) 0.48693426407352264 #'TFIDF'
(2, 3)  0.6176143709756019  #'Computation'
(2, 2)  0.6176143709756019  #'Calculation'
(3, 5)  0.2407515909314943  #'is'
(3, 10) 0.2973761110467491  #'TFIDF'
(3, 11) 0.37718388973255157 #'the'
(3, 7)  0.37718388973255157 #'product'
(3, 6)  0.37718388973255157 #'of'
(3, 9)  0.37718388973255157 #'TF'
(3, 0)  0.37718388973255157 #'and'
(3, 4)  0.37718388973255157 #'IDF'

因为它是规范的 TF-IDF 值,所以向量元素的平方和将等于 1。例如,对于索引 0 处的第一个文档,向量元素的平方和将等于 1:sum([0.6136667440107333 ** 2, 0.4968161174826459 ** 2, 0.6136667440107333 ** 2])

您可以通过设置来关闭此转换norm=None

print(TfidfVectorizer(norm=None).fit_transform(df[0])) #the same values you find in TfidfVectorizer(norm=None).fit_transform(df[0]).toarray(), but more readable
output: ([index of document on array lvl 0 / row], [index of unique word on array lvl 1 / column]) TF-IDF value
(0, 12) 1.5108256237659907 #1st word in 1st sentence: 'This'
(0, 5)  1.2231435513142097 #'is'
(0, 8)  1.5108256237659907 #'string', see that word 'a' is missing
(1, 12) 1.5108256237659907 #'This'
(1, 5)  1.2231435513142097 #'is'
(1, 8)  1.5108256237659907 #'string'
(1, 1)  1.916290731874155  #'another'
(2, 10) 1.5108256237659907 #'TFIDF'
(2, 3)  1.916290731874155  #'Computation'
(2, 2)  1.916290731874155  #'Calculation'
(3, 5)  1.2231435513142097 #'is'
(3, 10) 1.5108256237659907 #'TFIDF'
(3, 11) 1.916290731874155  #'the'
(3, 7)  1.916290731874155  #'product'
(3, 6)  1.916290731874155  #'of'
(3, 9)  1.916290731874155  #'TF'
(3, 0)  1.916290731874155  #'and'
(3, 4)  1.916290731874155  #'IDF'

因为每个单词在每个文档中只出现一次,所以 TF-IDF 值是每个单词的 IDF 值乘以 1:

tfidf_vectorizer = TfidfVectorizer(norm=None)
tfidf = tfidf_vectorizer.fit_transform(df[0])
print(tfidf_vectorizer.idf_)
output: Smoothed IDF-values
[1.91629073 1.91629073 1.91629073 1.91629073 1.91629073 1.22314355
 1.91629073 1.91629073 1.51082562 1.91629073 1.51082562 1.91629073
 1.51082562]

以上,希望对你有所帮助。

不幸的是,我无法重现转换,因为

当应用 l2 范数时,两个向量之间的余弦相似度是它们的点积。

似乎是一个额外的步骤。因为当您使用默认设置时,TF-IDF 值会因每个文档中的字数而有所偏差norm='l2',所以我会简单地通过使用来关闭此设置norm=None。我发现,您不能简单地使用以下方法进行转换:

tfidf_norm_calculated = [
    [(word/sum(doc))**0.5 for word in doc]
    for doc in TfidfVectorizer(norm=None).fit_transform(df[0]).toarray()]
print(tfidf_norm_calculated)
print('Sum of squares of vector elements is 1: ', [sum([word**2 for word in doc]) for doc in tfidf_norm_calculated])
print('Compare to:', TfidfVectorizer().fit_transform(df[0]).toarray())
于 2019-03-04T13:19:08.020 回答