I would like to extract unique combinations of letters within words using the scikit-learn TF-IDF vectorizer for an NLP problem. However, I'm not interested in individual letters, but letter combinations, so that, e.g. "the" should produce "th" and "he" but not "t", "h" or "e". My understanding is I should be able to use ngram_range. However, using ngram_range=(2,3) is still returning unigrams.
Example:
from sklearn.feature_extraction.text import TfidfVectorizer
examples = ['The cat on the mat',
'Fast and bulbous']
tfidf = TfidfVectorizer(max_features=None,
analyzer='char_wb',
ngram_range=(2, 3))
data=tfidf.fit_transform(examples)
print(pd.DataFrame(data=data.todense(),
index=examples,
columns = tfidf.get_feature_names_out()))
gives me the 2- and 3-gram results as expected but also unigrams (i.e. I don't want "a", "b", etc.):
a an b bu c \
The cat on the mat 0.000000 0.000000 0.000000 0.000000 0.139994
Fast and bulbous 0.181053 0.181053 0.181053 0.181053 0.000000
ca f fa m ma ... \
The cat on the mat 0.139994 0.000000 0.000000 0.139994 0.139994 ...
Fast and bulbous 0.000000 0.181053 0.181053 0.000000 0.000000 ...
s st st t th \
The cat on the mat 0.000000 0.000000 0.000000 0.199213 0.279987
Fast and bulbous 0.181053 0.181053 0.181053 0.128821 0.000000
the ul ulb us us
The cat on the mat 0.279987 0.000000 0.000000 0.000000 0.000000
Fast and bulbous 0.000000 0.181053 0.181053 0.181053 0.181053
[2 rows x 53 columns]
I would've expected this output with ngram_range=(1,3) but not ngram_range=(2,3).
Edit:
I just noticed that "a" is extracted from "Fast and bulbous", presumably as it occurs as " a", i.e. with a space before the "a", but not in "The cat on the mat" as the "a" in "cat" is surrounded by "c" and "t". Likewise, "u" is not extracted as there is no space surrounding it in either text.
It seems like TfidfVectorizer is extracting bigrams including spaces. Is there a way to turn this off? (I though using analyzer='char_wb' searched within words rather than across words).
I constructed a callable to pass to analyzer. It is stolen from based on the function from the source code of TfidfVectorizer used when analyzer is set to 'char_wb':
def char_wb_ngrams(text_document, ngram_range):
"""Callable for TfidfVectorizer analyzer, based on _char_wb_ngrams from TfidfVectorizer source code at
https://github.com/scikit-learn/scikit-learn/blob/f3f51f9b6/sklearn/feature_extraction/text.py"""
ngrams = []
min_n, max_n = ngram_range
for w in text_document.lower().split():
# This line in _char_wb_ngrams pads words with spaces and needs to be removed:
#w = " " + w + " "
w_len = len(w)
for n in range(min_n, max_n + 1):
offset = 0
ngrams.append(w[offset : offset + n])
while offset + n < w_len:
offset += 1
ngrams.append(w[offset : offset + n])
if offset == 0: # count a short word (w_len < n) only once
break
return ngrams
This works on the example data from above:
from functools import partial
tfidf_no_space = TfidfVectorizer(max_features=None,
analyzer=partial(char_wb_ngrams, ngram_range=(2,3)),
ngram_range=(2, 3))
data=tfidf_no_space.fit_transform(examples)
print(pd.DataFrame(data=data.todense(),
index=examples,
columns = tfidf_no_space.get_feature_names_out()))
which yields
an and as ast at \
The cat on the mat 0.000000 0.000000 0.000000 0.000000 0.436436
Fast and bulbous 0.229416 0.229416 0.229416 0.229416 0.000000
bo bou bu bul ca ... \
The cat on the mat 0.000000 0.000000 0.000000 0.000000 0.218218 ...
Fast and bulbous 0.229416 0.229416 0.229416 0.229416 0.000000 ...
nd on ou ous st \
The cat on the mat 0.000000 0.218218 0.000000 0.000000 0.000000
Fast and bulbous 0.229416 0.000000 0.229416 0.229416 0.229416
th the ul ulb us
The cat on the mat 0.436436 0.436436 0.000000 0.000000 0.000000
Fast and bulbous 0.000000 0.000000 0.229416 0.229416 0.229416
[2 rows x 28 columns]
I'm not sure this would work with punctuation, though. It would be good to have a version that strips punctuation and also doesn't require the call to partial (which fixes ngram_range in the function).
Related
I am trying to generate a few columns in a dataframe with datetime index based on a rule which references their own previous values. I have tried a for loop on the length of df as per below but looking for a cleaner solution if possible?
Because what I want to do in the end is get the stats of generated columns (C,D,E in below example) over a large number of A,B....
import pandas as pd
import numpy as np
df = pd.DataFrame(np.random.randn(30, 2), columns=list('AB'))
reset_level = 0.5
df['diff'] = df['A'].diff()
df['C'], df['D'], df['E'] = [0.0, 0.0, 0.0]
for i in range(1,len(df)):
if abs(df.iloc[i-1]['C'] + df.iloc[i]['diff']) > (reset_level):
df.iat[i,3] = 0.000
df.iat[i,4] = (df.iloc[i-1]['C'] + df.iloc[i]['diff'])
else:
df.iat[i,3] = (df.iloc[i-1]['C'] + df.iloc[i]['diff'])
df.iat[i,4] = 0.000
df.iat[i,5] = 0.5 * df.iloc[i]['D'] * df.iloc[i]['D']
Edit : Adding expected output below
A B diff C D E
0 -0.352725 1.429037 NaN 0.000000 0.000000 0.000000
1 -1.024418 -0.644302 -0.671693 0.000000 -0.671693 0.225585
2 0.401065 0.419555 1.425483 0.000000 1.425483 1.016001
3 -1.302484 0.724320 -1.703549 0.000000 -1.703549 1.451039
4 0.427035 0.835221 1.729518 0.000000 1.729518 1.495617
5 0.158694 -0.416741 -0.268340 -0.268340 0.000000 0.000000
6 0.921985 -0.490635 0.763291 0.494951 0.000000 0.000000
7 -0.835297 -1.036580 -1.757282 0.000000 -1.262331 0.796740
8 0.752060 -0.279206 1.587356 0.000000 1.587356 1.259850
9 1.795306 -1.554886 1.043246 0.000000 1.043246 0.544181
10 -0.405100 -0.361454 -2.200406 0.000000 -2.200406 2.420893
11 -0.253629 -0.627245 0.151471 0.151471 0.000000 0.000000
12 -0.820573 -0.212886 -0.566944 -0.415473 0.000000 0.000000
13 0.473439 2.532487 1.294012 0.000000 0.878539 0.385916
14 -1.395435 1.016338 -1.868875 0.000000 -1.868875 1.746346
15 -0.244269 -0.337820 1.151166 0.000000 1.151166 0.662592
16 -2.084977 -1.262249 -1.840708 0.000000 -1.840708 1.694103
17 0.666323 -1.696245 2.751300 0.000000 2.751300 3.784825
18 0.235207 -0.513903 -0.431115 -0.431115 0.000000 0.000000
19 1.386456 -0.149153 1.151249 0.000000 0.720134 0.259296
20 0.093456 -0.298154 -1.293000 0.000000 -1.293000 0.835925
21 0.690499 -1.687416 0.597043 0.000000 0.597043 0.178230
22 1.287530 -1.390260 0.597031 0.000000 0.597031 0.178223
23 1.828138 -0.288829 0.540608 0.000000 0.540608 0.146128
24 0.209666 -0.903385 -1.618472 0.000000 -1.618472 1.309727
25 -1.010678 0.615569 -1.220344 0.000000 -1.220344 0.744619
26 -1.799800 1.536332 -0.789122 0.000000 -0.789122 0.311357
27 0.611096 -1.033066 2.410896 0.000000 2.410896 2.906209
28 -0.532675 -0.091541 -1.143770 0.000000 -1.143770 0.654105
29 2.468137 -1.046117 3.000811 0.000000 3.000811 4.502435
I converted your for loop using a numpy array to hold the conditions and then np.where to replace the values according to your condition:
Define condition array
condition = np.abs(df.C.shift() + df["diff"]) > reset_level
Replace the values according to condition
df.iloc[:, 3] = np.where(condition, np.zeros((df.shape[0])), (df['C'].shift() + df['diff']))
df.iloc[:, 4] = np.where(~condition, np.zeros((df.shape[0])), (df['C'].shift() + df['diff']))
df.iloc[:, 5] = 0.5 * df['D'] * df['D']
Output:
A B diff C D E
0 -0.432513 -0.259526 NaN NaN 0.000000 0.000000
1 -1.120872 -1.572850 -0.688360 0.000000 NaN NaN
2 -0.917555 -2.251316 0.203317 0.203317 0.000000 0.000000
3 -1.869781 -1.284524 -0.952225 0.000000 -0.748908 0.280432
4 -2.041950 -0.091837 -0.172169 -0.172169 0.000000 0.000000
5 -0.142499 0.207746 1.899451 0.000000 1.727282 1.491751
6 1.432833 0.085211 1.575332 0.000000 1.575332 1.240835
7 -2.500191 -0.009907 -3.933025 0.000000 -3.933025 7.734341
8 0.154460 -1.859954 2.654651 0.000000 2.654651 3.523587
9 -0.565057 -0.516736 -0.719517 0.000000 -0.719517 0.258853
10 0.329845 0.127978 0.894902 0.000000 0.894902 0.400425
11 -0.920558 1.254617 -1.250402 0.000000 -1.250402 0.781753
12 -1.396913 0.262378 -0.476355 -0.476355 0.000000 0.000000
13 0.117336 -0.439932 1.514249 0.000000 1.037894 0.538612
14 -0.227066 2.565831 -0.344402 -0.344402 0.000000 0.000000
15 0.077750 0.195277 0.304816 0.304816 0.000000 0.000000
16 1.470611 -0.357213 1.392861 0.000000 1.697677 1.441053
17 -0.553844 0.339270 -2.024455 0.000000 -2.024455 2.049209
18 -0.259603 0.212839 0.294242 0.294242 0.000000 0.000000
19 0.605961 0.279599 0.865564 0.000000 1.159805 0.672574
20 -0.326706 -0.774350 -0.932667 0.000000 -0.932667 0.434934
21 -0.927601 -2.360751 -0.600895 0.000000 -0.600895 0.180537
22 -0.372085 0.986228 0.555516 0.000000 0.555516 0.154299
23 -0.687731 -2.966817 -0.315647 -0.315647 0.000000 0.000000
24 -0.041028 -0.328898 0.646703 0.000000 0.331057 0.054799
25 0.099489 0.275983 0.140517 0.140517 0.000000 0.000000
26 0.468274 -0.287097 0.368785 0.368785 0.000000 0.000000
27 0.497417 -0.588481 0.029143 0.029143 0.000000 0.000000
28 0.603178 2.243163 0.105761 0.105761 0.000000 0.000000
29 -0.643283 -1.051491 -1.246461 0.000000 -1.140700 0.650598
Is this what you were looking for, you didn't provide expected output.
Documentation:
np.where
Try this one (but don't iterate over all rows - it will do the whole column at once for you):
df["C_prev"] = df["C"].shift(1)
import pandas as pd
import numpy as np
df = pd.DataFrame(np.random.randn(30, 2), columns=list('AB'))
reset_level = 0.5
df['diff'] = df['A'].diff()
df['C'], df['D'], df['E'] = [0.0, 0.0, 0.0]
Then apply a function to each row:
def f(row):
if abs(df.loc[row.name - 1, 'C'] + row['diff']) > reset_level:
C = 0.0
D = df.loc[row.name - 1, 'C'] + row['diff']
else:
C = df.loc[row.name - 1, 'C'] + row['diff']
D = 0.0
E = 0.5 * row['D'] * row['D']
return(pd.Series([C, D, E]))
df.loc[1:, ['C', 'D', 'E']] = df[1:].apply(f, axis=1)
when I build a matrix using the last row of my dataframe:
x = w.iloc[-1, :]
a = np.mat(x).T
it goes:
ValueError: ndarray is not contiguous
`print the x shows(I have 61 columns in my dataframe):
print(x)
cdl2crows 0.000000
cdl3blackcrows 0.000000
cdl3inside 0.000000
cdl3linestrike 0.000000
cdl3outside 0.191465
cdl3starsinsouth 0.000000
cdl3whitesoldiers_x 0.000000
cdl3whitesoldiers_y 0.000000
cdladvanceblock 0.000000
cdlhighwave 0.233690
cdlhikkake 0.218209
cdlhikkakemod 0.000000
...
cdlidentical3crows 0.000000
cdlinneck 0.000000
cdlinvertedhammer 0.351235
cdlkicking 0.000000
cdlkickingbylength 0.000000
cdlladderbottom 0.002259
cdllongleggeddoji 0.629053
cdllongline 0.588480
cdlmarubozu 0.065362
cdlmatchinglow 0.032838
cdlmathold 0.000000
cdlmorningdojistar 0.000000
cdlmorningstar 0.327749
cdlonneck 0.000000
cdlpiercing 0.251690
cdlrickshawman 0.471466
cdlrisefall3methods 0.000000
Name: 2010-01-04, Length: 61, dtype: float64
how to solve it? so many thanks
np.mat expects array form of input.
refer to the doc
doc
So your code should be
x = w.iloc[-1, :].values
a = np.mat(x).T
.values will give numpy array format of dataframe values, so np.mat will work.
Use np.array instead of np.mat:
a = np.array(x).T
I need to calculate the tfidf matrix for few sentences. sentence include both numbers and words.
I am using below code to do so
import pandas as pd
from sklearn.feature_extraction.text import TfidfVectorizer
data1=['1/8 wire','4 tube','1-1/4 brush']
dataset=pd.DataFrame(data1, columns=['des'])
vectorizer1 = TfidfVectorizer(lowercase=False)
tf_idf_matrix = pd.DataFrame(vectorizer1.fit_transform(dataset['des']).toarray(),columns=vectorizer1.get_feature_names())
Tfidf function is considering only words as its vocabulary i.e
Out[3]: ['brush', 'tube', 'wire']
but i need numbers to be part of tokens
expected
Out[3]: ['brush', 'tube', 'wire','1/8','4','1-1/4']
After reading TfidfVectorizer documentation, I came to know have to change token_pattern and tokenizer parameters. But I am not getting how to change it to consider numbers and punctuation.
can anyone please tell me how to change the parameters.
You're right, token_pattern requires a custom regex pattern, pass a regex that treats any one or more characters that don't contain whitespace characters as a single token.
tfidf = TfidfVectorizer(lowercase=False, token_pattern=r'\S+')
tf_idf_matrix = pd.DataFrame(
tfidf.fit_transform(dataset['des']).toarray(),
columns=tfidf.get_feature_names()
)
print(tf_idf_matrix)
1-1/4 1/8 4 brush tube wire
0 0.000000 0.707107 0.000000 0.000000 0.000000 0.707107
1 0.000000 0.000000 0.707107 0.000000 0.707107 0.000000
2 0.707107 0.000000 0.000000 0.707107 0.000000 0.000000
you can explicitly point out in token_pattern parameter the symbols you would like to parse:
token_pattern_ = r'([a-zA-Z0-9-/]{1,})'
where {1,} indicates the minimum number of symbols the word should contain. End then you pass this as a parameter to token_pattern:
tfidf = TfidfVectorizer(token_pattern = token_pattern_)
Hi I have a lemmatized text in the format as shown by lemma. I want to get TfIdf score for each word this is the function that I wrote:
import numpy as np
import pandas as pd
from sklearn.feature_extraction.text import TfidfVectorizer
lemma=["'Ah", 'yes', u'say', 'softly', 'Harry',
'Potter', 'Our', 'new', 'celebrity', 'You',
'learn', 'subtle', 'science', 'exact', 'art',
'potion-making', u'begin', 'He', u'speak', 'barely',
'whisper', 'caught', 'every', 'word', 'like',
'Professor', 'McGonagall', 'Snape', 'gift',
u'keep', 'class', 'silent', 'without', 'effort',
'As', 'little', 'foolish', 'wand-waving', 'many',
'hardly', 'believe', 'magic', 'I', 'dont', 'expect', 'really',
'understand', 'beauty']
def Tfidf_Vectorize(lemmas_name):
vect = TfidfVectorizer(stop_words='english',ngram_range=(1,2))
vect_transform = vect.fit_transform(lemmas_name)
# First approach of creating a dataframe of weight & feature names
vect_score = np.asarray(vect_transform.mean(axis=0)).ravel().tolist()
vect_array = pd.DataFrame({'term': vect.get_feature_names(), 'weight': vect_score})
vect_array.sort_values(by='weight',ascending=False,inplace=True)
# Second approach of getting the feature names
vect_fn = np.array(vect.get_feature_names())
sorted_tfidf_index = vect_transform.max(0).toarray()[0].argsort()
print('Largest Tfidf:\n{}\n'.format(vect_fn[sorted_tfidf_index[:-11:-1]]))
return vect_array
tf_dataframe=Tfidf_Vectorize(lemma)
print(tf_dataframe.iloc[:5,:])
The output I am getting by:
print('Largest Tfidf:\n{}\n'.format(vect_fn[sorted_tfidf_index[:-11:-1]]))
is
Largest Tfidf:
[u'yes' u'fools' u'fury' u'gale' u'ghosts' u'gift' u'glory' u'glow' u'good'
u'granger']
The result of tf_dataframe
term weight
261 snape 0.027875
238 say 0.022648
211 potter 0.013937
181 mind 0.010453
123 harry 0.010453
60 dark 0.006969
75 dumbledore 0.006969
311 voice 0.005226
125 head 0.005226
231 ron 0.005226
Shouldn't both approaches lead to the same result of top features? I just want to calculate the tfidf scores and get the top 5 features/weight. What am i doing wrong?
I am not sure what I am looking at here but I have the feeling that you're using TfidfVectorizer incorrectly. However, please correct me in case I got the wrong idea of what you're trying.
So.. what you need is a list of documents which you feed to fit_transform(). From that you can construct a matrix where, for example, each column represents a document and each row a word. One cell in that matrix is the tf-idf score of the word i in document j.
Here's an example:
import numpy as np
import pandas as pd
from sklearn.feature_extraction.text import TfidfVectorizer
documents = [
"This is a document.",
"This is another document with slightly more text.",
"Whereas this is yet another document with even more text than the other ones.",
"This document is awesome and also rather long.",
"The car he drove was red."
]
document_names = ['Doc {:d}'.format(i) for i in range(len(documents))]
def get_tfidf(docs, ngram_range=(1,1), index=None):
vect = TfidfVectorizer(stop_words='english', ngram_range=ngram_range)
tfidf = vect.fit_transform(documents).todense()
return pd.DataFrame(tfidf, columns=vect.get_feature_names(), index=index).T
print(get_tfidf(documents, ngram_range=(1,2), index=document_names))
Which will give you:
Doc 0 Doc 1 Doc 2 Doc 3 Doc 4
awesome 0.0 0.000000 0.000000 0.481270 0.000000
awesome long 0.0 0.000000 0.000000 0.481270 0.000000
car 0.0 0.000000 0.000000 0.000000 0.447214
car drove 0.0 0.000000 0.000000 0.000000 0.447214
document 1.0 0.282814 0.282814 0.271139 0.000000
document awesome 0.0 0.000000 0.000000 0.481270 0.000000
document slightly 0.0 0.501992 0.000000 0.000000 0.000000
document text 0.0 0.000000 0.501992 0.000000 0.000000
drove 0.0 0.000000 0.000000 0.000000 0.447214
drove red 0.0 0.000000 0.000000 0.000000 0.447214
long 0.0 0.000000 0.000000 0.481270 0.000000
ones 0.0 0.000000 0.501992 0.000000 0.000000
red 0.0 0.000000 0.000000 0.000000 0.447214
slightly 0.0 0.501992 0.000000 0.000000 0.000000
slightly text 0.0 0.501992 0.000000 0.000000 0.000000
text 0.0 0.405004 0.405004 0.000000 0.000000
text ones 0.0 0.000000 0.501992 0.000000 0.000000
The two methods you show to get to words and their respective scores calculate the mean over all documents and fetch the max score of each word respectively.
So let's do this and compare the two methods:
df = get_tfidf(documents, ngram_range=(1,2), index=index)
print(pd.DataFrame([df.mean(1), df.max(1)], index=['score_mean', 'score_max']).T)
We can see that the scores are of course different.
score_mean score_max
awesome 0.096254 0.481270
awesome long 0.096254 0.481270
car 0.089443 0.447214
car drove 0.089443 0.447214
document 0.367353 1.000000
document awesome 0.096254 0.481270
document slightly 0.100398 0.501992
document text 0.100398 0.501992
drove 0.089443 0.447214
drove red 0.089443 0.447214
long 0.096254 0.481270
ones 0.100398 0.501992
red 0.089443 0.447214
slightly 0.100398 0.501992
slightly text 0.100398 0.501992
text 0.162002 0.405004
text ones 0.100398 0.501992
Note:
You can convince yourself that this does the same as calling min/max on the TfidfVectorizer:
vect = TfidfVectorizer(stop_words='english', ngram_range=(1,2))
tfidf = vect.fit_transform(documents)
print(tfidf.max(0))
print(tfidf.mean(0))
I am trying to use nipype to analyze transformation matrixes that were created by FSL.
FSL has a script called "avscale" that analyzes those transformation matrixes (*.mat files).
I was wondering whether nipype has any interface that wrap that script and enable to work with its output.
Thanks
Based on the docs and the current source the answer is no. Also, avscale has also not been mentioned on the nipy-devel mailing list since at least last February. It's possible that Nipype already wraps something else that does this (perhaps with a matlab wrapper?) You could try opening an issue or asking the the mailing list.
As long as you're trying to use Python (with nipype and all), maybe the philosophy of the nipype project is that you should just use numpy/scipy for this? Just a guess, I don't know the functions to replicate this output with those tools. It's also possible that no one has gotten around to adding it yet.
For the uninitiated, avscale takes this affine matrix:
1.00614 -8.39414e-06 0 -0.757356
0 1.00511 -0.00317841 -0.412038
0 0.0019063 1.00735 -0.953364
0 0 0 1
and yields this or similar output:
Rotation & Translation Matrix:
1.000000 0.000000 0.000000 -0.757356
0.000000 0.999998 -0.001897 -0.412038
0.000000 0.001897 0.999998 -0.953364
0.000000 0.000000 0.000000 1.000000
Scales (x,y,z) = 1.006140 1.005112 1.007354
Skews (xy,xz,yz) = -0.000008 0.000000 -0.001259
Average scaling = 1.0062
Determinant = 1.01872
Left-Right orientation: preserved
Forward half transform =
1.003065 -0.000004 -0.000000 -0.378099
0.000000 1.002552 -0.001583 -0.206133
0.000000 0.000951 1.003669 -0.475711
0.000000 0.000000 0.000000 1.000000
Backward half transform =
0.996944 0.000004 0.000000 0.376944
0.000000 0.997452 0.001575 0.206357
0.000000 -0.000944 0.996343 0.473777
0.000000 0.000000 0.000000 1.000000