Iterate nltk.tokenize across all rows of a pandas dataframe

Question

grateful for your help for what feels like a stupid question. I've pulled a sqlite table into a pandas dataframe so I can tokenize and count the frequency of words from a series of tweets.

With the code below, I can produce this for the first tweet. How do I iterate for the whole table?

conn = sqlite3.connect("tweets.sqlite")
data = pd.read_sql_query("select tweet_text from tweets_new;", conn)

tokenizer=RegexpTokenizer(r'\w+')
tokens=tokenizer.tokenize(data['tweet_text'][0])

words = nltk.FreqDist(tokens)

unigram_df = pd.DataFrame(words.most_common(),
                             columns=["WORD","COUNT"])

unigram_df

When I change the value to anything other than a single row, I get the following error:

TypeError: expected string or buffer

I know there are other ways of doing this, but I need to do it along these lines because of how I intend to use the output next. Thanks for any help you can provide!

I have tried:

%%time

tokenizer = RegexpTokenizer(r'\w+')  

print "Cleaning the tweets...\n"
for i in xrange(0,len(df)):
    if( (i+1)%1000000 == 0 ):  
        tokens=tokenizer.tokenize(df['tweet_text'][i])
        words = nltk.FreqDist(tokens)

This looks like it should work, but still only returns words from the first row.

Answer 1

In case anyone is interested in this niche use case, here's the code I was eventually able to make work:

conn = sqlite3.connect("tweets.sqlite")
data = pd.read_sql_query("select tweet_text from tweets_new;", conn)

alldata = str(data)

tokenizer=RegexpTokenizer(r'\w+')
tokens=tokenizer.tokenize(alldata)

words = nltk.FreqDist(tokens)

unigram_df = pd.DataFrame(words.most_common(),
                             columns=["WORD","COUNT"])

Thanks for your help everyone!

Answer 2

I think your problem can be solved more concisely using CountVectorizer. I'll give you an example. Given the following inputs:

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

corpus_tweets = [['I love pizza and hambuerger'],['I love apple and chips'], ['The pen is on the table!!']]
df = pd.DataFrame(corpus_tweets, columns=['tweet_text'])

You can create your bag of words template with these few lines:

count_vect = CountVectorizer()
X_train_counts = count_vect.fit_transform(df.tweet_text)

You can print the obtained vocabulary:

count_vect.vocabulary_
# ouutput: {'love': 5, 'pizza': 8, 'and': 0, 'hambuerger': 3, 'apple': 1, 'chips': 2, 'the': 10, 'pen': 7, 'is': 4, 'on': 6, 'table': 9}

and get the dataframe with word counts:

df_count = pd.DataFrame(X_train_counts.todense(), columns=count_vect.get_feature_names())

   and  apple  chips  hambuerger  is  love  on  pen  pizza  table  the
0    1      0      0           1   0     1   0    0      1      0    0
1    1      1      1           0   0     1   0    0      0      0    0
2    0      0      0           0   1     0   1    1      0      1    2

If it is useful for you, you can merge the dataframe of the counts with the dataframe of the corpus:

pd.concat([df, df_count],  axis=1)

                    tweet_text  and  apple  chips  hambuerger  is  love  on  \
0  I love pizza and hambuerger    1      0      0           1   0     1   0   
1       I love apple and chips    1      1      1           0   0     1   0   
2    The pen is on the table!!    0      0      0           0   1     0   1   

   pen  pizza  table  the  
0    0      1      0    0  
1    0      0      0    0  
2    1      0      1    2  

If you want to get the dictionary containing the <word, count> pairs for each document, at this point all you need to do is:

dict_count = df_count.T.to_dict()

{0: {'and': 1,
  'apple': 0,
  'chips': 0,
  'hambuerger': 1,
  'is': 0,
  'love': 1,
  'on': 0,
  'pen': 0,
  'pizza': 1,
  'table': 0,
  'the': 0},
 1: {'and': 1,
  'apple': 1,
  'chips': 1,
  'hambuerger': 0,
  'is': 0,
  'love': 1,
  'on': 0,
  'pen': 0,
  'pizza': 0,
  'table': 0,
  'the': 0},
 2: {'and': 0,
  'apple': 0,
  'chips': 0,
  'hambuerger': 0,
  'is': 1,
  'love': 0,
  'on': 1,
  'pen': 1,
  'pizza': 0,
  'table': 1,
  'the': 2}}

Note: turning X_train_counts which is a sparse numpy matrix into a dataframe is not a good idea. But it can be useful to understand and visualize the various steps of your model.

Answer 3

After creating the DataFrame loop over all the rows:

tokenizer = RegexpTokenizer(r'\w+')
fdist = FreqDist()
for txt in data['tweet_text']:
      for word in tokenizer.tokenize(txt):
          fdist[word.lower()] += 1