How can I make Scikit-learn TfidfVectorizer not to preprocess the text?

Question

I reading some text data from a csv and trying to build a TF-IDF feature vector using those data.

The data looks something like:

where the content contains specially formatted strings (synset).

When I try to build a TF-IDF vector with that, I am expecting to preserve that format, but when I do

tfidf = TfidfVectorizer()
data['content'] = data['content'].fillna('')
tfidf_matrix = tfidf.fit_transform(data['content'])

and look at the tfidf.vocabulary_

The text data is preprocessed as:

{'square': 3754,
 '01': 0,
 '02': 1,
 'public_square': 3137,
 '04': 3,
 '05': 4,
 '06': 5,
 '07': 6,
 '08': 7,
 '03': 2,
 'feather': 1666,
 'straight': 3821,...

I want it to count square.n.01 as a single text instead of splitting it up.

I would be able to do this if I build TF-IDF from scratch, but I feel like that is unnecessary. Any help?

Answer 1

you need to write your own tokenization function which need to be called in tokenizer parameter of tfidfVectorizer

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


df = pd.DataFrame(data = [[['square.n.01','square.n.02','public_square.n.01']], 
[['two.n.01','deuce.n.04','two.s.01']]], columns = ['content'])

df['content'] = df['content'].astype(str)
df['content'] = df['content'].apply(lambda x: x.replace('[','').replace(']',''))

def my_tokenizer(doc):
    return doc.split(',')

tfidf = TfidfVectorizer(tokenizer = my_tokenizer)
tfidf_matrix = tfidf.fit_transform(df['content'])

print(tfidf.vocabulary_)
#o/p
{"'square.n.01'": 4,
 " 'square.n.02'": 2,
 " 'public_square.n.01'": 1,
 "'two.n.01'": 5,
 " 'deuce.n.04'": 0,
 " 'two.s.01'": 3}