这是正确的tfidf吗?

时间:2017-08-14 18:05:51

标签: python scikit-learn tf-idf

我想从文档中获取tfidf。但我不认为它给了我正确的价值观,或者我可能做错了什么。请建议。代码和输出如下:

from sklearn.feature_extraction.text import TfidfVectorizer
books = ["Hello there this is first book to be read by wordcount script.", "This is second book to be read by wordcount script. It has some additionl information.", "just third book."]
vectorizer = TfidfVectorizer()
response = vectorizer.fit_transform(books)
feature_names = vectorizer.get_feature_names()
for col in response.nonzero()[1]:
   print feature_names[col], '-', response[0, col]

更新1 :(如juanpa.arrivillaga所建议)

vectorizer = TfidfVectorizer(smooth_idf=False)

输出:

script - 0.269290317245
wordcount - 0.269290317245
by - 0.269290317245
read - 0.269290317245
be - 0.269290317245
to - 0.269290317245
book - 0.209127954024
first - 0.354084405732
is - 0.269290317245
this - 0.269290317245
there - 0.354084405732
hello - 0.354084405732
information - 0.0
...

更新1后的输出:

script - 0.256536760895
wordcount - 0.256536760895
by - 0.256536760895
read - 0.256536760895
be - 0.256536760895
to - 0.256536760895
book - 0.182528018244
first - 0.383055542114
is - 0.256536760895
this - 0.256536760895
there - 0.383055542114
hello - 0.383055542114
information - 0.0
...

根据我的理解,tfidf是= tf * idf。我手动计算它的方式如下:

文件1:“你好,这是第一本由wordcount脚本阅读的书。” 文档2:“这是wordcount脚本读取的第二本书。它有一些附加信息。” 文件3:“只是第三本书。”

Tfidf for Hello:

tf= 1/12(total terms in document 1)= 0.08333333333
idf= log(3(total documents)/1(no. of document with term in it))= 0.47712125472
0.08333333333*0.47712125472= 0.03976008865 

与以下不同(hello - 0.354084405732)。

更新1后的手动计算:

tf = 1
idf= log(nd/df) +1 = log (3/1) +1= 0.47712125472 + 1= 1.47712 
tfidf = tf*idf = 1* 1.47712= 1.47712

(与idf平滑后的代码输出“hello - 0.383055542114”不同)

任何有助于了解最新情况的帮助都非常感谢..

1 个答案:

答案 0 :(得分:3)

这是没有平滑或标准化的输出:

In [2]: from sklearn.feature_extraction.text import TfidfVectorizer
   ...: books = ["Hello there this is first book to be read by wordcount script.", "This is second book to be read by wordcount sc
   ...: ript. It has some additionl information.", "just third book."]
   ...: vectorizer = TfidfVectorizer(smooth_idf=False, norm=None)
   ...: response = vectorizer.fit_transform(books)
   ...: feature_names = vectorizer.get_feature_names()
   ...: for col in response.nonzero()[1]:
   ...:    print(feature_names[col], '-', response[0, col])
   ...:
hello - 2.09861228867
there - 2.09861228867
this - 1.40546510811
is - 1.40546510811
first - 2.09861228867
book - 1.0
to - 1.40546510811
be - 1.40546510811
read - 1.40546510811
by - 1.40546510811
wordcount - 1.40546510811
script - 1.40546510811
this - 1.40546510811
is - 1.40546510811
book - 1.0
to - 1.40546510811
be - 1.40546510811
read - 1.40546510811
by - 1.40546510811
wordcount - 1.40546510811
script - 1.40546510811
second - 0.0
it - 0.0
has - 0.0
some - 0.0
additionl - 0.0
information - 0.0
book - 1.0
just - 0.0
third - 0.0

请考虑"hello"的结果:

hello - 2.09861228867

现在,手动:

In [3]: import math

In [4]: tf = 1

In [5]: idf = math.log(3/1) + 1

In [6]: tf*idf
Out[6]: 2.09861228866811

手动计算的问题在于您使用的是log base 10 ,但您需要使用自然对数。

如果你仍然觉得要经历平滑和标准化步骤的强烈愿望,那么这应该让你正确地做到这一点。