Skip to content

Instantly share code, notes, and snippets.

@bbengfort
Created October 26, 2014 21:12
Show Gist options
  • Save bbengfort/9646279d220994222845 to your computer and use it in GitHub Desktop.
Save bbengfort/9646279d220994222845 to your computer and use it in GitHub Desktop.
Uses TFIDF to extract relevant sentences from text. Based off of Charlie Greenbacker's example from "A smattering of NLP with Python" presentation he gave a while ago.
# summarize
# Uses TFIDF to extract relevent sentences from text.
#
# Author: Benjamin Bengfort <[email protected]>
# Created: Sun Oct 26 16:06:36 2014 -0400
#
# ID: summarize.py [] [email protected] $
"""
Uses TFIDF to extract relevent sentences from text.
Based off of Charlie Greenbacker's example from "A smattering of NLP with
Python" presentation he gave a while ago.
"""
##########################################################################
## Imports
##########################################################################
from __future__ import division
import re
import sys
import math
import nltk
import time
import random
from operator import itemgetter
from sklearn.feature_extraction.text import TfidfVectorizer
##########################################################################
## Module Fixtures
##########################################################################
lemmatizer = nltk.stem.WordNetLemmatizer()
stopwords = set(nltk.corpus.stopwords.words('english'))
def timeit(func):
def wrapper(*args, **kwargs):
start = time.time()
result = func(*args, **kwargs)
finit = time.time()
return result, finit-start
return wrapper
def tokenize(text):
for token in nltk.word_tokenize(text):
token = token.lower()
if token in stopwords or not token.isalpha():
continue
yield lemmatizer.lemmatize(token)
class TfidfSummarizer(object):
def __init__(self, corpus, verbose=2):
self.corpus = corpus
self.tokenizer = tokenize
self.vectorizer = TfidfVectorizer(tokenizer=self.tokenizer, decode_error='ignore')
self.tdm = None
self.features = None
self.initialize_features(verbose)
def initialize_features(self, verbose=1):
"""
Initializes the features and prints time taken if verbose > 0.
"""
@timeit
def inner(self):
# articles = dict((fileid, self.corpus.raw(fileid)) for fileid in self.corpus.fileids())
return self.vectorizer.fit_transform(self.corpus.raw(fileid) for fileid in self.corpus.fileids())
self.tdm, delta = inner(self)
self.features = self.vectorizer.get_feature_names()
if verbose > 0:
print "TDM contains %i terms and %i documents." % (len(self.features), self.tdm.shape[0])
if verbose > 1:
print " First term: %s" % self.features[0]
print " Last term: %s" % self.features[-1]
for idx in xrange(0, 4):
print " Random term: %s" % random.choice(self.features[1:-2])
if verbose > 0:
print "Featurization took %0.3f seconds\n" % delta
def score_sentences(self, fileid):
fileidx = self.corpus.fileids().index(fileid)
for sent in nltk.sent_tokenize(self.corpus.raw(fileid)):
score = 0
for token in self.tokenizer(sent):
if token not in self.features: continue
score += self.tdm[fileidx, self.features.index(token)]
yield score, sent
def summarize(self, fileid):
"""
Returns a summary for the given fileid
"""
scores = list(self.score_sentences(fileid))
sumlen = int(math.ceil(len(scores) / 2 ))
scores.sort(key=itemgetter(0))
return scores[:sumlen]
if __name__ == '__main__':
summarizer = TfidfSummarizer(nltk.corpus.reuters)
article = random.choice(nltk.corpus.reuters.fileids())
print "#"*75
print "## SUMMARY"
print "#"*75
print
for sent in summarizer.summarize(article):
print sent[1]
print
print "#"*75
print "## ORIGINAL"
print "#"*75
print nltk.corpus.reuters.raw(article)
print
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment