Fast and Effective ways to Extract Keyphrases using TFIDF with Python

import os       
path = "/content/drive/MyDrive/Data/theses100/theses100/"    
all_files = os.listdir(path+"docsutf8")
all_keys = os.listdir(path+ "keys")
print(len(all_files)," files n",all_files, "n", all_keys)  # won't necessarily be sorted

all_documents =[]
all_keys = []
all_files_names = []
for i, fname in enumerate(all_files):
  with open(path+'docsutf8/'+fname) as f:
      lines = f.readlines()
  key_name= fname[:-4]
  with open(path+'keys/'+key_name+'.key') as f:
      k = f.readlines()
  all_text = ' '.join(lines)
  keyss = ' '.join(k)
  all_documents.append(all_text)
  all_keys.append(keyss.split("n"))
  all_files_names.append(key_name)

#------------------------------------------------------------------------------

Now we have the data that needs to be cleaned up to remove unwanted noise.

Text Pre-Processing

Preprocessing includes tokenization, lemmatization, lowercasing, removing numbers, white spaces and words shorter than three letters, removing stop words, remove symbols and punctuation. We are not going to explain these functions because that is not the purpose of this article.

For lemmatization, WordNetLemmatizer was used. I do not recommend performing the steaming because it changes the root of the word in some cases.

#clean text applying all the text preprocessing functions
dtf['cleaned_text'] = dtf.text.apply(lambda x: ' '.join(preprocess_text(x)))
dtf.head()

However, before using this line of code you must define the function preprocess_text first. Make sure to write the following code and then apply the preprocess_text function to the dataframe.

Importing the used libraries

import nltknltk.download('stopwords')
nltk.download('wordnet')

# For cleaning the text

import spacy

import nltk

from nltk.tokenize import word_tokenize

from nltk.corpus import stopwords

import regex as re

import string

import numpy as np

import nltk.data

import re

nltk.download('punkt')

nltk.download('averaged_perceptron_tagger')

from nltk.stem import WordNetLemmatizer

from nltk import word_tokenize, sent_tokenize, pos_tag

def preprocess_text(text):

# 1. Tokenise to alphabetic tokens

text = remove_numbers(text)

text = remove_http(text)

text = remove_punctuation(text)

text = convert_to_lower(text)

text = remove_white_space(text)

text = remove_short_words(text)

tokens = toknizing(text)

# 2. POS tagging

pos_map = {'J': 'a', 'N': 'n', 'R': 'r', 'V': 'v'}

pos_tags_list = pos_tag(tokens)

#print(pos_tags)

# 3. Lowercase and lemmatise

lemmatiser = WordNetLemmatizer()

tokens = [lemmatiser.lemmatize(w.lower(), pos=pos_map.get(p[0], 'v')) for w, p in pos_tags_list]

return tokens

#------------------------------------------------------------------------------

def convert_to_lower(text):

return text.lower()

#------------------------------------------------------------------------------

def remove_numbers(text):

text = re.sub(r'd+' , '', text)

return text

#------------------------------------------------------------------------------

def remove_http(text):

text = re.sub("https?://t.co/[A-Za-z0-9]*", ' ', text)

return text

#------------------------------------------------------------------------------

def remove_short_words(text):

text = re.sub(r'bw{1,2}b', '', text)

return text

#------------------------------------------------------------------------------

def remove_punctuation(text):

punctuations = '''!()[]{};«№»:'",`./?@=#$-(%^)+&[*_]~'''

no_punctuation = ""




for char in text:

if char not in punctuations:

no_punctuation = no_punctuation + char

return no_punctuation

#------------------------------------------------------------------------------

def remove_white_space(text):

text = text.strip()

return text

#------------------------------------------------------------------------------

def toknizing(text):

stop_words = set(stopwords.words('english'))

tokens = word_tokenize(text)

## Remove Stopwords from tokens

result = [i for i in tokens if not i in stop_words]

return result

After that, we clean the golden keywords for each document and perform lemmatization to facilitate the matching process later with the words that will be produced by the TFIDF with Python algorithm.

# Clean the basic keywords and remove the spaces and noise
def clean_orginal_kw(orginal_kw):
  orginal_kw_clean =[]
  for doc_kw in orginal_kw:
    temp =[]
    for t in doc_kw:
      tt = ' '.join(preprocess_text(t))
      if len(tt.split())>0:
        temp.append(tt)
    orginal_kw_clean.append(temp)
  return orginal_kw_clean
#_______________________________________________
orginal_kw= clean_orginal_kw(dtf['goldkeys'])

TFIDF Keywords Extraction

1. Generating n-grams (keyphrases) and weighing them

First we import Tfidf Vectorizer from the text feature extraction package. In the second line we set idf=true i.e. we want to use the inverse document frequency IDF with the term frequency. Its maximum value is 0.5, which means that we only want terms that occur in 50 per cent of the documents (for our case in 49 documents, out of 99). If a term appears in more than 50 documents, it will be deleted because it is considered non-discriminatory at the corpus level. We specify the range of n-grams from one to three (you can set it to a larger number, but according to the statistics of the current data set, the largest proportion is for keywords that are 1-3 in length)

The third line generates documents’ vectors.

from sklearn.feature_extraction.text import TfidfVectorizer
vectorizer = TfidfVectorizer(use_idf=True, max_df=0.5,min_df=1, ngram_range=(1,3))
vectors = vectorizer.fit_transform(dtf['cleaned_text'])

Then for each document, we want to build a dictionary (dict_of_tokens) where the key is the word and the value is the TFIDF weight. We create a list of tfidf_vectors to store the dictionaries of all documents.

dict_of_tokens={i[1]:i[0] for i in vectorizer.vocabulary_.items()}

tfidf_vectors = []  # all deoc vectors by tfidf
for row in vectors:
  tfidf_vectors.append({dict_of_tokens[column]:value for (column,value) in zip(row.indices,row.data)})

Let’s see what this dictionary contains for the first document

The content of the dictionary for the first document

Everything makes sense! The number of dictionaries is the same as the number of documents, and we see that the dictionary of the first document contains every n-gram with its TFIDFweight.

2. Sorting keyphrases by TFIDF weights

The next step is to simply sort the n-grams in each dictionary according to the TFIDF weight in descending order. Set “reverse=True” to make the order descend.

doc_sorted_tfidfs =[]  # list of doc features each with tfidf weight
#sort each dict of a document
for dn in tfidf_vectors:
  newD = sorted(dn.items(), key=lambda x: x[1], reverse=True)
  newD = dict(newD)
  doc_sorted_tfidfs.append(newD)

Now you can get a list of keywords without their weights.

For example, let’s choose the Top 5 keywords for the first document.

Great, it works!

Performance Evaluation

The above is sufficient to use the method in extracting keywords or keyphrases, but in the following, we want to evaluate the effectiveness of the method in a scientific way according to a standard metric for this type of task.

First I would like to note that we use an exact match for the evaluation, where the keyphrase automatically extracted from the document must exactly match the document’s gold standard keyphrase.

Keyword extraction is a ranking problem. One of the most commonly used measures for ranking is “Mean average precision at K, MAP@K. To calculate MAP@K, the precision at K elements p@k is first considered as the basic metric of the ranking quality for one document.

def apk(kw_actual, kw_predicted, top_k=10):
    if len(kw_predicted)>top_k:
        kw_predicted = kw_predicted[:top_k]
    score = 0.0
    num_hits = 0.0
    for i,p in enumerate(kw_predicted):
        if p in kw_actual and p not in kw_predicted[:i]:
            num_hits += 1.0
            score += num_hits / (i+1.0)
    if not kw_actual:
        return 0.0
    return score / min(len(kw_actual), top_k)
def mapk(kw_actual, kw_predicted, top_k=10):
    return np.mean([apk(a,p,top_k) for a,p in zip(kw_actual, kw_predicted)])

This function apk accepts two parameters: the list of keywords predicted by the TFIDF method (kw_predicted), and the list of gold standard keywords (kw_actual). The default value for k is 10. Here we print MAP value at k=[5,10,20,40].

for k in [5, 10,20,40]:
  mpak= mapk(orginal_kw, tfidf_kw, k)
  print("mean average precession  @",k,  '=  {0:.4g}'.format(mpak))

Conclusion

In this article, we have presented an easy and simple way to extract keywords from documents using TFIDF with Python. The code was written in Python and explained step by step. The performance of the method was evaluated using the MAP criterion as a ranking task. This method, despite its simplicity, is very effective and is considered one of the strong baselines in this field. I hope the content was useful for you. You can check the code on my repository at GitHub. I would be grateful for any feedback.

All images are screenshots by me (Author: Ali Mansour)

Want to know more about TFIDF with Python? Read here.

Ali Mansour 31 Dec 2021