Python: tf-idf-cosine: to find document similarity
I was following a tutorial which was available at Part 1 & Part 2. Unfortunately the author didn't have the time for the final section which involved using cosine similarity to actually find the distance between two documents. I followed the examples in the article with the help of the following link from stackoverflow, included is the code mentioned in the above link (just so as to make life easier)
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfTransformer
from nltk.corpus import stopwords
import numpy as np
import numpy.linalg as LA
train_set = ["The sky is blue.", "The sun is bright."] # Documents
test_set = ["The sun in the sky is bright."] # Query
stopWords = stopwords.words('english')
vectorizer = CountVectorizer(stop_words = stopWords)
#print vectorizer
transformer = TfidfTransformer()
#print transformer
trainVectorizerArray = vectorizer.fit_transform(train_set).toarray()
testVectorizerArray = vectorizer.transform(test_set).toarray()
print 'Fit Vectorizer to train set', trainVectorizerArray
print 'Transform Vectorizer to test set', testVectorizerArray
transformer.fit(trainVectorizerArray)
print
print transformer.transform(trainVectorizerArray).toarray()
transformer.fit(testVectorizerArray)
print
tfidf = transformer.transform(testVectorizerArray)
print tfidf.todense()
as a result of the above code I have the following matrix
Fit Vectorizer to train set [[1 0 1 0]
[0 1 0 1]]
Transform Vectorizer to test set [[0 1 1 1]]
[[ 0.70710678 0. 0.70710678 0. ]
[ 0. 0.70710678 0. 0.70710678]]
[[ 0. 0.57735027 0.57735027 0.57735027]]
I am not sure how to use this output in order to calculate cosine similarity, I know how to implement cosine similarity with respect to two vectors of similar length but here I am not sure how to identify the two vectors.
Solution 1:
First off, if you want to extract count features and apply TF-IDF normalization and row-wise euclidean normalization you can do it in one operation with TfidfVectorizer:
>>> from sklearn.feature_extraction.text import TfidfVectorizer
>>> from sklearn.datasets import fetch_20newsgroups
>>> twenty = fetch_20newsgroups()
>>> tfidf = TfidfVectorizer().fit_transform(twenty.data)
>>> tfidf
<11314x130088 sparse matrix of type '<type 'numpy.float64'>'
with 1787553 stored elements in Compressed Sparse Row format>
Now to find the cosine distances of one document (e.g. the first in the dataset) and all of the others you just need to compute the dot products of the first vector with all of the others as the tfidf vectors are already row-normalized.
As explained by Chris Clark in comments and here Cosine Similarity does not take into account the magnitude of the vectors. Row-normalised have a magnitude of 1 and so the Linear Kernel is sufficient to calculate the similarity values.
The scipy sparse matrix API is a bit weird (not as flexible as dense N-dimensional numpy arrays). To get the first vector you need to slice the matrix row-wise to get a submatrix with a single row:
>>> tfidf[0:1]
<1x130088 sparse matrix of type '<type 'numpy.float64'>'
with 89 stored elements in Compressed Sparse Row format>
scikit-learn already provides pairwise metrics (a.k.a. kernels in machine learning parlance) that work for both dense and sparse representations of vector collections. In this case we need a dot product that is also known as the linear kernel:
>>> from sklearn.metrics.pairwise import linear_kernel
>>> cosine_similarities = linear_kernel(tfidf[0:1], tfidf).flatten()
>>> cosine_similarities
array([ 1. , 0.04405952, 0.11016969, ..., 0.04433602,
0.04457106, 0.03293218])
Hence to find the top 5 related documents, we can use argsort and some negative array slicing (most related documents have highest cosine similarity values, hence at the end of the sorted indices array):
>>> related_docs_indices = cosine_similarities.argsort()[:-5:-1]
>>> related_docs_indices
array([ 0, 958, 10576, 3277])
>>> cosine_similarities[related_docs_indices]
array([ 1. , 0.54967926, 0.32902194, 0.2825788 ])
The first result is a sanity check: we find the query document as the most similar document with a cosine similarity score of 1 which has the following text:
>>> print twenty.data[0]
From: [email protected] (where's my thing)
Subject: WHAT car is this!?
Nntp-Posting-Host: rac3.wam.umd.edu
Organization: University of Maryland, College Park
Lines: 15
I was wondering if anyone out there could enlighten me on this car I saw
the other day. It was a 2-door sports car, looked to be from the late 60s/
early 70s. It was called a Bricklin. The doors were really small. In addition,
the front bumper was separate from the rest of the body. This is
all I know. If anyone can tellme a model name, engine specs, years
of production, where this car is made, history, or whatever info you
have on this funky looking car, please e-mail.
Thanks,
- IL
---- brought to you by your neighborhood Lerxst ----
The second most similar document is a reply that quotes the original message hence has many common words:
>>> print twenty.data[958]
From: [email protected] (Robert Seymour)
Subject: Re: WHAT car is this!?
Article-I.D.: reed.1993Apr21.032905.29286
Reply-To: [email protected]
Organization: Reed College, Portland, OR
Lines: 26
In article <[email protected]> [email protected] (where's my
thing) writes:
>
> I was wondering if anyone out there could enlighten me on this car I saw
> the other day. It was a 2-door sports car, looked to be from the late 60s/
> early 70s. It was called a Bricklin. The doors were really small. In
addition,
> the front bumper was separate from the rest of the body. This is
> all I know. If anyone can tellme a model name, engine specs, years
> of production, where this car is made, history, or whatever info you
> have on this funky looking car, please e-mail.
Bricklins were manufactured in the 70s with engines from Ford. They are rather
odd looking with the encased front bumper. There aren't a lot of them around,
but Hemmings (Motor News) ususally has ten or so listed. Basically, they are a
performance Ford with new styling slapped on top.
> ---- brought to you by your neighborhood Lerxst ----
Rush fan?
--
Robert Seymour [email protected]
Physics and Philosophy, Reed College (NeXTmail accepted)
Artificial Life Project Reed College
Reed Solar Energy Project (SolTrain) Portland, OR
Solution 2:
WIth the Help of @excray's comment, I manage to figure it out the answer, What we need to do is actually write a simple for loop to iterate over the two arrays that represent the train data and test data.
First implement a simple lambda function to hold formula for the cosine calculation:
cosine_function = lambda a, b : round(np.inner(a, b)/(LA.norm(a)*LA.norm(b)), 3)
And then just write a simple for loop to iterate over the to vector, logic is for every "For each vector in trainVectorizerArray, you have to find the cosine similarity with the vector in testVectorizerArray."
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.feature_extraction.text import TfidfTransformer
from nltk.corpus import stopwords
import numpy as np
import numpy.linalg as LA
train_set = ["The sky is blue.", "The sun is bright."] #Documents
test_set = ["The sun in the sky is bright."] #Query
stopWords = stopwords.words('english')
vectorizer = CountVectorizer(stop_words = stopWords)
#print vectorizer
transformer = TfidfTransformer()
#print transformer
trainVectorizerArray = vectorizer.fit_transform(train_set).toarray()
testVectorizerArray = vectorizer.transform(test_set).toarray()
print 'Fit Vectorizer to train set', trainVectorizerArray
print 'Transform Vectorizer to test set', testVectorizerArray
cx = lambda a, b : round(np.inner(a, b)/(LA.norm(a)*LA.norm(b)), 3)
for vector in trainVectorizerArray:
print vector
for testV in testVectorizerArray:
print testV
cosine = cx(vector, testV)
print cosine
transformer.fit(trainVectorizerArray)
print
print transformer.transform(trainVectorizerArray).toarray()
transformer.fit(testVectorizerArray)
print
tfidf = transformer.transform(testVectorizerArray)
print tfidf.todense()
Here is the output:
Fit Vectorizer to train set [[1 0 1 0]
[0 1 0 1]]
Transform Vectorizer to test set [[0 1 1 1]]
[1 0 1 0]
[0 1 1 1]
0.408
[0 1 0 1]
[0 1 1 1]
0.816
[[ 0.70710678 0. 0.70710678 0. ]
[ 0. 0.70710678 0. 0.70710678]]
[[ 0. 0.57735027 0.57735027 0.57735027]]
Solution 3:
I know its an old post. but I tried the http://scikit-learn.sourceforge.net/stable/ package. here is my code to find the cosine similarity. The question was how will you calculate the cosine similarity with this package and here is my code for that
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.metrics.pairwise import cosine_similarity
from sklearn.feature_extraction.text import TfidfVectorizer
f = open("/root/Myfolder/scoringDocuments/doc1")
doc1 = str.decode(f.read(), "UTF-8", "ignore")
f = open("/root/Myfolder/scoringDocuments/doc2")
doc2 = str.decode(f.read(), "UTF-8", "ignore")
f = open("/root/Myfolder/scoringDocuments/doc3")
doc3 = str.decode(f.read(), "UTF-8", "ignore")
train_set = ["president of India",doc1, doc2, doc3]
tfidf_vectorizer = TfidfVectorizer()
tfidf_matrix_train = tfidf_vectorizer.fit_transform(train_set) #finds the tfidf score with normalization
print "cosine scores ==> ",cosine_similarity(tfidf_matrix_train[0:1], tfidf_matrix_train) #here the first element of tfidf_matrix_train is matched with other three elements
Here suppose the query is the first element of train_set and doc1,doc2 and doc3 are the documents which I want to rank with the help of cosine similarity. then I can use this code.
Also the tutorials provided in the question was very useful. Here are all the parts for it part-I,part-II,part-III
the output will be as follows :
[[ 1. 0.07102631 0.02731343 0.06348799]]
here 1 represents that query is matched with itself and the other three are the scores for matching the query with the respective documents.
Solution 4:
Let me give you another tutorial written by me. It answers your question, but also makes an explanation why we are doing some of the things. I also tried to make it concise.
So you have a list_of_documents which is just an array of strings and another document which is just a string. You need to find such document from the list_of_documents that is the most similar to document.
Let's combine them together: documents = list_of_documents + [document]
Let's start with dependencies. It will become clear why we use each of them.
from nltk.corpus import stopwords
import string
from nltk.tokenize import wordpunct_tokenize as tokenize
from nltk.stem.porter import PorterStemmer
from sklearn.feature_extraction.text import TfidfVectorizer
from scipy.spatial.distance import cosine
One of the approaches that can be uses is a bag-of-words approach, where we treat each word in the document independent of others and just throw all of them together in the big bag. From one point of view, it looses a lot of information (like how the words are connected), but from another point of view it makes the model simple.
In English and in any other human language there are a lot of "useless" words like 'a', 'the', 'in' which are so common that they do not possess a lot of meaning. They are called stop words and it is a good idea to remove them. Another thing that one can notice is that words like 'analyze', 'analyzer', 'analysis' are really similar. They have a common root and all can be converted to just one word. This process is called stemming and there exist different stemmers which differ in speed, aggressiveness and so on. So we transform each of the documents to list of stems of words without stop words. Also we discard all the punctuation.
porter = PorterStemmer()
stop_words = set(stopwords.words('english'))
modified_arr = [[porter.stem(i.lower()) for i in tokenize(d.translate(None, string.punctuation)) if i.lower() not in stop_words] for d in documents]
So how will this bag of words help us? Imagine we have 3 bags: [a, b, c], [a, c, a] and [b, c, d]. We can convert them to vectors in the basis [a, b, c, d]. So we end up with vectors: [1, 1, 1, 0], [2, 0, 1, 0] and [0, 1, 1, 1]. The similar thing is with our documents (only the vectors will be way to longer). Now we see that we removed a lot of words and stemmed other also to decrease the dimensions of the vectors. Here there is just interesting observation. Longer documents will have way more positive elements than shorter, that's why it is nice to normalize the vector. This is called term frequency TF, people also used additional information about how often the word is used in other documents - inverse document frequency IDF. Together we have a metric TF-IDF which have a couple of flavors. This can be achieved with one line in sklearn :-)
modified_doc = [' '.join(i) for i in modified_arr] # this is only to convert our list of lists to list of strings that vectorizer uses.
tf_idf = TfidfVectorizer().fit_transform(modified_doc)
Actually vectorizer allows to do a lot of things like removing stop words and lowercasing. I have done them in a separate step only because sklearn does not have non-english stopwords, but nltk has.
So we have all the vectors calculated. The last step is to find which one is the most similar to the last one. There are various ways to achieve that, one of them is Euclidean distance which is not so great for the reason discussed here. Another approach is cosine similarity. We iterate all the documents and calculating cosine similarity between the document and the last one:
l = len(documents) - 1
for i in xrange(l):
minimum = (1, None)
minimum = min((cosine(tf_idf[i].todense(), tf_idf[l + 1].todense()), i), minimum)
print minimum
Now minimum will have information about the best document and its score.