def train(): positive_tweets = read_tweets('/root/295/new/positive.txt', 'positive') negative_tweets = read_tweets('/root/295/new/negative.txt', 'negative') print len(positive_tweets) print len(negative_tweets) #pos_train = positive_tweets[:2000] #neg_train = negative_tweets[:2000] #pos_test = positive_tweets[2001:3000] #neg_test = negative_tweets[2001:3000] pos_train = positive_tweets[:len(positive_tweets)*80/100] neg_train = negative_tweets[:len(negative_tweets)*80/100] pos_test = positive_tweets[len(positive_tweets)*80/100+1:] neg_test = negative_tweets[len(positive_tweets)*80/100+1:] training_data = pos_train + neg_train test_data = pos_test + neg_test sentim_analyzer = SentimentAnalyzer() all_words_neg = sentim_analyzer.all_words([mark_negation(doc) for doc in training_data]) #print all_words_neg unigram_feats = sentim_analyzer.unigram_word_feats(all_words_neg, min_freq=4) #print unigram_feats print len(unigram_feats) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) training_set = sentim_analyzer.apply_features(training_data) test_set = sentim_analyzer.apply_features(test_data) print test_set trainer = NaiveBayesClassifier.train classifier = sentim_analyzer.train(trainer, training_set) for key,value in sorted(sentim_analyzer.evaluate(test_set).items()): print('{0}: {1}'.format(key, value)) print sentim_analyzer.classify(tokenize_sentance('I hate driving car at night')) return sentim_analyzer
def sentiment_analysis(self, testing_data, training_data=None): if training_data is None: training_data = self.training_data ## Apply sentiment analysis to data to extract new "features" # Initialize sentiment analyzer object sentiment_analyzer = SentimentAnalyzer() # Mark all negative words in training data, using existing list of negative words all_negative_words = sentiment_analyzer.all_words([mark_negation(data) for data in training_data]) unigram_features = sentiment_analyzer.unigram_word_feats(all_negative_words, min_freq=4) len(unigram_features) sentiment_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_features) training_final = sentiment_analyzer.apply_features(training_data) testing_final = sentiment_analyzer.apply_features(testing_data) ## Traing model and test model = NaiveBayesClassifier.train classifer = sentiment_analyzer.train(model, training_final) for key, value in sorted(sentiment_analyzer.evaluate(testing_final).items()): print ("{0}: {1}".format(key, value))
def demo_subjectivity(trainer, save_analyzer=False, n_instances=None, output=None): """ Train and test a classifier on instances of the Subjective Dataset by Pang and Lee. The dataset is made of 5000 subjective and 5000 objective sentences. All tokens (words and punctuation marks) are separated by a whitespace, so we use the basic WhitespaceTokenizer to parse the data. :param trainer: `train` method of a classifier. :param save_analyzer: if `True`, store the SentimentAnalyzer in a pickle file. :param n_instances: the number of total sentences that have to be used for training and testing. Sentences will be equally split between positive and negative. :param output: the output file where results have to be reported. """ from nltk.sentiment import SentimentAnalyzer from nltk.corpus import subjectivity if n_instances is not None: n_instances = int(n_instances/2) subj_docs = [(sent, 'subj') for sent in subjectivity.sents(categories='subj')[:n_instances]] obj_docs = [(sent, 'obj') for sent in subjectivity.sents(categories='obj')[:n_instances]] # We separately split subjective and objective instances to keep a balanced # uniform class distribution in both train and test sets. train_subj_docs, test_subj_docs = split_train_test(subj_docs) train_obj_docs, test_obj_docs = split_train_test(obj_docs) training_docs = train_subj_docs+train_obj_docs testing_docs = test_subj_docs+test_obj_docs sentim_analyzer = SentimentAnalyzer() all_words_neg = sentim_analyzer.all_words([mark_negation(doc) for doc in training_docs]) # Add simple unigram word features handling negation unigram_feats = sentim_analyzer.unigram_word_feats(all_words_neg, min_freq=4) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) # Apply features to obtain a feature-value representation of our datasets training_set = sentim_analyzer.apply_features(training_docs) test_set = sentim_analyzer.apply_features(testing_docs) classifier = sentim_analyzer.train(trainer, training_set) try: classifier.show_most_informative_features() except AttributeError: print('Your classifier does not provide a show_most_informative_features() method.') results = sentim_analyzer.evaluate(test_set) if save_analyzer == True: save_file(sentim_analyzer, 'sa_subjectivity.pickle') if output: extr = [f.__name__ for f in sentim_analyzer.feat_extractors] output_markdown(output, Dataset='subjectivity', Classifier=type(classifier).__name__, Tokenizer='WhitespaceTokenizer', Feats=extr, Instances=n_instances, Results=results) return sentim_analyzer
def demo_movie_reviews(trainer, n_instances=None, output=None): """ Train classifier on all instances of the Movie Reviews dataset. The corpus has been preprocessed using the default sentence tokenizer and WordPunctTokenizer. Features are composed of: - most frequent unigrams :param trainer: `train` method of a classifier. :param n_instances: the number of total reviews that have to be used for training and testing. Reviews will be equally split between positive and negative. :param output: the output file where results have to be reported. """ from nltk.corpus import movie_reviews from nltk.sentiment import SentimentAnalyzer if n_instances is not None: n_instances = int(n_instances/2) pos_docs = [(list(movie_reviews.words(pos_id)), 'pos') for pos_id in movie_reviews.fileids('pos')[:n_instances]] neg_docs = [(list(movie_reviews.words(neg_id)), 'neg') for neg_id in movie_reviews.fileids('neg')[:n_instances]] # We separately split positive and negative instances to keep a balanced # uniform class distribution in both train and test sets. train_pos_docs, test_pos_docs = split_train_test(pos_docs) train_neg_docs, test_neg_docs = split_train_test(neg_docs) training_docs = train_pos_docs+train_neg_docs testing_docs = test_pos_docs+test_neg_docs sentim_analyzer = SentimentAnalyzer() all_words = sentim_analyzer.all_words(training_docs) # Add simple unigram word features unigram_feats = sentim_analyzer.unigram_word_feats(all_words, min_freq=4) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) # Apply features to obtain a feature-value representation of our datasets training_set = sentim_analyzer.apply_features(training_docs) test_set = sentim_analyzer.apply_features(testing_docs) classifier = sentim_analyzer.train(trainer, training_set) try: classifier.show_most_informative_features() except AttributeError: print('Your classifier does not provide a show_most_informative_features() method.') results = sentim_analyzer.evaluate(test_set) if output: extr = [f.__name__ for f in sentim_analyzer.feat_extractors] output_markdown(output, Dataset='Movie_reviews', Classifier=type(classifier).__name__, Tokenizer='WordPunctTokenizer', Feats=extr, Results=results, Instances=n_instances)
def train_model(training): ## Apply sentiment analysis to data to extract new "features" # Initialize sentiment analyzer object sentiment_analyzer = SentimentAnalyzer() # Mark all negative words in training data, using existing list of negative words all_negative_words = sentiment_analyzer.all_words([mark_negation(data) for data in training]) unigram_features = sentiment_analyzer.unigram_word_feats(all_negative_words, min_freq=4) len(unigram_features) sentiment_analyzer.add_feat_extractor(extract_unigram_feats,unigrams=unigram_features) training_final = sentiment_analyzer.apply_features(training) return [training_final]
def get_objectivity_analyzer(): n_instances = 100 subj_docs = [(sent, 'subj') for sent in subjectivity.sents(categories='subj')[:n_instances]] obj_docs = [(sent, 'obj') for sent in subjectivity.sents(categories='obj')[:n_instances]] train_subj_docs = subj_docs train_obj_docs = obj_docs training_docs = train_subj_docs+train_obj_docs sentim_analyzer = SentimentAnalyzer() all_words_neg = sentim_analyzer.all_words([mark_negation(doc) for doc in training_docs]) unigram_feats = sentim_analyzer.unigram_word_feats(all_words_neg, min_freq=4) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) training_set = sentim_analyzer.apply_features(training_docs) trainer = NaiveBayesClassifier.train sentiment_classifier = sentim_analyzer.train(trainer, training_set) return sentim_analyzer
(['smart', 'and', 'alert', ',', 'thirteen', 'conversations', 'about', 'one', 'thing', 'is', 'a', 'small', 'gem', '.'], 'subj') train_subj_docs = subj_docs[:80] test_subj_docs = subj_docs[80:100] train_obj_docs = obj_docs[:80] test_obj_docs = obj_docs[80:100] training_docs = train_subj_docs+train_obj_docs testing_docs = test_subj_docs+test_obj_docs sentim_analyzer = SentimentAnalyzer() all_words_neg = sentim_analyzer.all_words([mark_negation(doc) for doc in training_docs]) unigram_feats = sentim_analyzer.unigram_word_feats(all_words_neg, min_freq=1) len(unigram_feats) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) training_set = sentim_analyzer.apply_features(training_docs) test_set = sentim_analyzer.apply_features(testing_docs) trainer = NaiveBayesClassifier.train classifier = sentim_analyzer.train(trainer, training_set) for key,value in sorted(sentim_analyzer.evaluate(test_set).items()): print('{0}: {1}'.format(key, value)) from nltk.sentiment.vader import SentimentIntensityAnalyzer sentences = ["VADER is smart, handsome, and funny.", # positive sentence example "VADER is smart, handsome, and funny!", # punctuation emphasis handled correctly (sentiment intensity adjusted) "VADER is very smart, handsome, and funny.", # booster words handled correctly (sentiment intensity adjusted) "VADER is VERY SMART, handsome, and FUNNY.", # emphasis for ALLCAPS handled "VADER is VERY SMART, handsome, and FUNNY!!!",# combination of signals - VADER appropriately adjusts intensity "VADER is VERY SMART, really handsome, and INCREDIBLY FUNNY!!!",# booster words & punctuation make this close to ceiling for score
def demo_tweets(trainer, n_instances=None, output=None): """ Train and test Naive Bayes classifier on 10000 tweets, tokenized using TweetTokenizer. Features are composed of: - 1000 most frequent unigrams - 100 top bigrams (using BigramAssocMeasures.pmi) :param trainer: `train` method of a classifier. :param n_instances: the number of total tweets that have to be used for training and testing. Tweets will be equally split between positive and negative. :param output: the output file where results have to be reported. """ from nltk.tokenize import TweetTokenizer from nltk.sentiment import SentimentAnalyzer from nltk.corpus import twitter_samples, stopwords # Different customizations for the TweetTokenizer tokenizer = TweetTokenizer(preserve_case=False) # tokenizer = TweetTokenizer(preserve_case=True, strip_handles=True) # tokenizer = TweetTokenizer(reduce_len=True, strip_handles=True) if n_instances is not None: n_instances = int(n_instances/2) fields = ['id', 'text'] positive_json = twitter_samples.abspath("positive_tweets.json") positive_csv = 'positive_tweets.csv' json2csv_preprocess(positive_json, positive_csv, fields, limit=n_instances) negative_json = twitter_samples.abspath("negative_tweets.json") negative_csv = 'negative_tweets.csv' json2csv_preprocess(negative_json, negative_csv, fields, limit=n_instances) neg_docs = parse_tweets_set(negative_csv, label='neg', word_tokenizer=tokenizer) pos_docs = parse_tweets_set(positive_csv, label='pos', word_tokenizer=tokenizer) # We separately split subjective and objective instances to keep a balanced # uniform class distribution in both train and test sets. train_pos_docs, test_pos_docs = split_train_test(pos_docs) train_neg_docs, test_neg_docs = split_train_test(neg_docs) training_tweets = train_pos_docs+train_neg_docs testing_tweets = test_pos_docs+test_neg_docs sentim_analyzer = SentimentAnalyzer() # stopwords = stopwords.words('english') # all_words = [word for word in sentim_analyzer.all_words(training_tweets) if word.lower() not in stopwords] all_words = [word for word in sentim_analyzer.all_words(training_tweets)] # Add simple unigram word features unigram_feats = sentim_analyzer.unigram_word_feats(all_words, top_n=1000) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) # Add bigram collocation features bigram_collocs_feats = sentim_analyzer.bigram_collocation_feats([tweet[0] for tweet in training_tweets], top_n=100, min_freq=12) sentim_analyzer.add_feat_extractor(extract_bigram_feats, bigrams=bigram_collocs_feats) training_set = sentim_analyzer.apply_features(training_tweets) test_set = sentim_analyzer.apply_features(testing_tweets) classifier = sentim_analyzer.train(trainer, training_set) # classifier = sentim_analyzer.train(trainer, training_set, max_iter=4) try: classifier.show_most_informative_features() except AttributeError: print('Your classifier does not provide a show_most_informative_features() method.') results = sentim_analyzer.evaluate(test_set) if output: extr = [f.__name__ for f in sentim_analyzer.feat_extractors] output_markdown(output, Dataset='labeled_tweets', Classifier=type(classifier).__name__, Tokenizer=tokenizer.__class__.__name__, Feats=extr, Results=results, Instances=n_instances)
import nltk from nltk.tokenize import word_tokenize from nltk.classify import NaiveBayesClassifier from nltk.sentiment import SentimentAnalyzer from nltk.sentiment.util import * f = open("training_set.txt",'r') sa = SentimentAnalyzer() trainingset = [] for line in f: senti = line.split(",")[0] content = line[len(senti)+1:] tokens = word_tokenize(content.rstrip()) trainingset.append((tokens,senti)) all_words_neg = sa.all_words([mark_negation(doc) for doc in trainingset]) unigram_feats = sa.unigram_word_feats(all_words_neg,min_freq = 4) sa.add_feat_extractor(extract_unigram_feats,unigrams=unigram_feats) training_set = sa.apply_features(trainingset) for line in sys.stdin: if "username" in line: continue tweetWords=[] tweet= line.split(";")[4] likes = line.split(";")[3] likes = int(likes) if likes==0: num=1 else: num = 1+likes
def demo_movie_reviews(trainer, n_instances=None, output=None): """ Train classifier on all instances of the Movie Reviews dataset. The corpus has been preprocessed using the default sentence tokenizer and WordPunctTokenizer. Features are composed of: - most frequent unigrams :param trainer: `train` method of a classifier. :param n_instances: the number of total reviews that have to be used for training and testing. Reviews will be equally split between positive and negative. :param output: the output file where results have to be reported. """ from nltk.corpus import movie_reviews from nltk.sentiment import SentimentAnalyzer if n_instances is not None: n_instances = int(n_instances / 2) pos_docs = [(list(movie_reviews.words(pos_id)), 'pos') for pos_id in movie_reviews.fileids('pos')[:n_instances]] neg_docs = [(list(movie_reviews.words(neg_id)), 'neg') for neg_id in movie_reviews.fileids('neg')[:n_instances]] # We separately split positive and negative instances to keep a balanced # uniform class distribution in both train and test sets. train_pos_docs, test_pos_docs = split_train_test(pos_docs) train_neg_docs, test_neg_docs = split_train_test(neg_docs) training_docs = train_pos_docs + train_neg_docs testing_docs = test_pos_docs + test_neg_docs sentim_analyzer = SentimentAnalyzer() all_words = sentim_analyzer.all_words(training_docs) # Add simple unigram word features unigram_feats = sentim_analyzer.unigram_word_feats(all_words, min_freq=4) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) # Apply features to obtain a feature-value representation of our datasets training_set = sentim_analyzer.apply_features(training_docs) test_set = sentim_analyzer.apply_features(testing_docs) classifier = sentim_analyzer.train(trainer, training_set) try: classifier.show_most_informative_features() except AttributeError: print( 'Your classifier does not provide a show_most_informative_features() method.' ) results = sentim_analyzer.evaluate(test_set) if output: extr = [f.__name__ for f in sentim_analyzer.feat_extractors] output_markdown(output, Dataset='Movie_reviews', Classifier=type(classifier).__name__, Tokenizer='WordPunctTokenizer', Feats=extr, Results=results, Instances=n_instances)
class SuicideClassifier(object): def __init__(self, sentiment_only, num_phrases_to_track=20): # neg_phrases = filter_negative_phrases(load_csv_sentences('thoughtsandfeelings.csv')) # pos_phrases = filter_positive_phrases(load_csv_sentences('spiritualforums.csv')) # file_pos = open("pos_phrases.txt", 'w') # file_neg = open("neg_phrases.txt", 'w') # for item in pos_phrases: # print>>file_pos, item # for item in neg_phrases: # print>>file_neg, item self.recent_sentiment_scores = [] neg_file = open("ALL_neg_phrases_filtered.txt", "r") pos_file = open("webtext_phrases_with_lots_of_words.txt", "r") neg_phrases = neg_file.readlines() pos_phrases = pos_file.readlines() neg_docs = [] pos_docs = [] for phrase in neg_phrases: neg_docs.append((phrase.split(), 'suicidal')) for phrase in pos_phrases[:len(neg_phrases)]: pos_docs.append((phrase.split(), 'alright')) print len(neg_docs) print len(pos_docs) # negcutoff = len(neg_docs) * 3 / 4 # poscutoff = len(pos_docs) * 3 / 4 negcutoff = -200 poscutoff = -200 train_pos_docs = pos_docs[:poscutoff] test_pos_docs = pos_docs[poscutoff:] train_neg_docs = neg_docs[:negcutoff] test_neg_docs = neg_docs[negcutoff:] training_docs = train_pos_docs + train_neg_docs testing_docs = test_pos_docs + test_neg_docs self.sentim_analyzer = SentimentAnalyzer() if not sentiment_only: all_words = self.sentim_analyzer.all_words([doc for doc in training_docs]) unigram_feats = self.sentim_analyzer.unigram_word_feats(all_words, min_freq=1) self.sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) self.sentim_analyzer.add_feat_extractor(vader_sentiment_feat) # bigram_feats = self.sentim_analyzer.bigram_collocation_feats(all_words, min_freq=1) # self.sentim_analyzer.add_feat_extractor(extract_bigram_feats, bigrams=bigram_feats) training_set = self.sentim_analyzer.apply_features(training_docs) test_set = self.sentim_analyzer.apply_features(testing_docs) trainer = NaiveBayesClassifier.train self.classifier = self.sentim_analyzer.train(trainer, training_set) for key, value in sorted(self.sentim_analyzer.evaluate(test_set).items()): print('{0}: {1}'.format(key, value)) self.classifier.show_most_informative_features(20) def test(self, phrase): return self.sentim_analyzer.classify(phrase.split()) def update_sentiments(self, value): now = datetime.datetime.now() self.recent_sentiment_scores.append([now, value]) self.recent_sentiment_scores = [x for x in self.recent_sentiment_scores if x[ 0] > now - datetime.timedelta(seconds=60)] print sum([x[1] for x in self.recent_sentiment_scores]) / len(self.recent_sentiment_scores) return sum([x[1] for x in self.recent_sentiment_scores]) / len(self.recent_sentiment_scores)
def demo_tweets(trainer, n_instances=None, output=None): """ Train and test Naive Bayes classifier on 10000 tweets, tokenized using TweetTokenizer. Features are composed of: - 1000 most frequent unigrams - 100 top bigrams (using BigramAssocMeasures.pmi) :param trainer: `train` method of a classifier. :param n_instances: the number of total tweets that have to be used for training and testing. Tweets will be equally split between positive and negative. :param output: the output file where results have to be reported. """ from nltk.tokenize import TweetTokenizer from nltk.sentiment import SentimentAnalyzer from nltk.corpus import twitter_samples, stopwords # Different customizations for the TweetTokenizer tokenizer = TweetTokenizer(preserve_case=False) # tokenizer = TweetTokenizer(preserve_case=True, strip_handles=True) # tokenizer = TweetTokenizer(reduce_len=True, strip_handles=True) if n_instances is not None: n_instances = int(n_instances / 2) fields = ['id', 'text'] positive_json = twitter_samples.abspath("positive_tweets.json") positive_csv = 'positive_tweets.csv' json2csv_preprocess(positive_json, positive_csv, fields, limit=n_instances) negative_json = twitter_samples.abspath("negative_tweets.json") negative_csv = 'negative_tweets.csv' json2csv_preprocess(negative_json, negative_csv, fields, limit=n_instances) neg_docs = parse_tweets_set(negative_csv, label='neg', word_tokenizer=tokenizer) pos_docs = parse_tweets_set(positive_csv, label='pos', word_tokenizer=tokenizer) # We separately split subjective and objective instances to keep a balanced # uniform class distribution in both train and test sets. train_pos_docs, test_pos_docs = split_train_test(pos_docs) train_neg_docs, test_neg_docs = split_train_test(neg_docs) training_tweets = train_pos_docs + train_neg_docs testing_tweets = test_pos_docs + test_neg_docs sentim_analyzer = SentimentAnalyzer() # stopwords = stopwords.words('english') # all_words = [word for word in sentim_analyzer.all_words(training_tweets) if word.lower() not in stopwords] all_words = [word for word in sentim_analyzer.all_words(training_tweets)] # Add simple unigram word features unigram_feats = sentim_analyzer.unigram_word_feats(all_words, top_n=1000) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) # Add bigram collocation features bigram_collocs_feats = sentim_analyzer.bigram_collocation_feats( [tweet[0] for tweet in training_tweets], top_n=100, min_freq=12) sentim_analyzer.add_feat_extractor(extract_bigram_feats, bigrams=bigram_collocs_feats) training_set = sentim_analyzer.apply_features(training_tweets) test_set = sentim_analyzer.apply_features(testing_tweets) classifier = sentim_analyzer.train(trainer, training_set) # classifier = sentim_analyzer.train(trainer, training_set, max_iter=4) try: classifier.show_most_informative_features() except AttributeError: print( 'Your classifier does not provide a show_most_informative_features() method.' ) results = sentim_analyzer.evaluate(test_set) if output: extr = [f.__name__ for f in sentim_analyzer.feat_extractors] output_markdown(output, Dataset='labeled_tweets', Classifier=type(classifier).__name__, Tokenizer=tokenizer.__class__.__name__, Feats=extr, Results=results, Instances=n_instances)
"I like the movie .".split()) # ['I', 'like', 'the', 'movie.'] print mark_negation("I don't like the movie .".split() ) # ['I', "don't", 'like_NEG', 'the_NEG', 'movie._NEG'] # The nltk classifier won't be able to handle the whole training set TRAINING_COUNT = 5000 analyzer = SentimentAnalyzer() vader = SentimentIntensityAnalyzer() vocabulary = analyzer.all_words([ mark_negation(word_tokenize(unidecode(clean_text(instance)))) for instance in train_X[:TRAINING_COUNT] ]) print "Vocabulary: ", len(vocabulary) # 1356908 print "Computing Unigran Features ..." unigram_features = analyzer.unigram_word_feats(vocabulary, min_freq=10) print "Unigram Features: ", len(unigram_features) # 8237 analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_features) def vader_polarity(text): """ Transform the output to a binary 0/1 result """ score = vader.polarity_scores(text) ''' if score['neu'] > ((score['pos'] + score['neg'])* 6.0): point = 'Neutral' elif score['pos'] > score['neg']: point = 'Positive' else: point = 'Negative' return point '''
subj_docs[0] train_subj_docs = subj_docs[:80] test_subj_docs = subj_docs[80:100] train_obj_docs = obj_docs[:80] test_obj_docs = obj_docs[80:100] training_docs = train_subj_docs + train_obj_docs testing_docs = test_subj_docs + test_obj_docs sentim_analyzer = SentimentAnalyzer() all_words_neg = sentim_analyzer.all_words( [mark_negation(doc) for doc in training_docs]) unigram_feats = sentim_analyzer.unigram_word_feats(all_words_neg, min_freq=4) len(unigram_feats) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) training_set = sentim_analyzer.apply_features(training_docs) test_set = sentim_analyzer.apply_features(testing_docs) trainer = NaiveBayesClassifier.train classifier = sentim_analyzer.train(trainer, training_set) for key, value in sorted(sentim_analyzer.evaluate(test_set).items()): print('{0}: {1}'.format(key, value)) ''' sentences = ["VADER is smart, handsome, and funny.1", "VADER is smart, handsome, and funny!", "VADER is very smart, handsome, and funny.", "VADER is VERY SMART, handsome, and FUNNY.", "VADER is VERY SMART, handsome, and FUNNY!!!", "VADER is VERY SMART, really handsome, and INCREDIBLY FUNNY!!!",
class QuoteFinder: def __init__(self): self.sentim_analyzer = SentimentAnalyzer() self.genre_dict = read_file("jsons/movie_genre_quote_dict_2.json") context_file = "jsons/final_context.json" movie_file = "jsons/final_movies.json" quote_file = "jsons/final_quotes.json" year_rating_file = "jsons/final_year_rating.json" self.context = read_file(context_file) self.movies = read_file(movie_file) self.quotes = read_file(quote_file) self.year_rating_dict = read_file(year_rating_file) # Reincode to unicode for i in range(len(self.context)): self.context[i] = self.context[i].encode("utf-8").decode("utf-8") self.movies[i] = self.movies[i].encode("utf-8").decode("utf-8") self.quotes[i] = self.quotes[i].encode("utf-8").decode("utf-8") self.context, self.quotes, self.movies = quote_pruner(self.context, self.quotes, self.movies) self.inverted_index = read_file("jsons/f_inverted_index.json") self.idf = read_file("jsons/f_idf.json") # Initialize query tokenizer self.tokenizer = TreebankWordTokenizer() # Compute document norms self.norms = compute_doc_norms(self.inverted_index, self.idf, len(self.context)) word_co_filename = "jsons/word_co.json" word_count_filename = "jsons/word_count_dict.json" pmi_dict_filename = "jsons/pmi_dict.json" # Read files self.word_co = read_file(word_co_filename) self.word_count_dict = read_file(word_count_filename) self.pmi_dict = read_file(pmi_dict_filename) def find_basic_cooccurence(self, word_list): """ Initialize the base word co-occurrance list from our context and quotes. Arguments ========= word_list: the list of words which are in our movie space Returns ======= word_co : a dictionary representing the word_occurrance matrix """ # Get English stop words stop_words = stopwords.words('english') # Merge context and quotes quote_list = self.quotes new_quote_list = [] for q in quote_list: new_q = punct_strip(q) if new_q not in self.context: new_quote_list.append(new_q) context_quotes = self.context + new_quote_list # Find co occurences in context data, based co-occurences in a document word_co = defaultdict(list) word_count_dict = defaultdict(int) for doc in context_quotes: # Double loop to count word co-occurences tkns = self.tokenizer.tokenize(doc) for i in range(len(tkns)): if tkns[i] not in stop_words: word_count_dict[tkns[i]] += 1 for j in range(len(tkns)): if not (j == i) and (tkns[j] in word_list): word_co[tkns[i]] = update_word_counts(word_co[tkns[i]], tkns[j]) return word_co, word_count_dict def update_cooccurence(self, word_co_old, word_count_dict_old, word_list, docs): """ Updates the word co-occurrance mat and the word count dict with a new set of data. Arguments ========= word_co_old: a word co-occurrance matrix in the form of a dictionary word_count_dict_old: a dictionary that keeps track of the total occurences of a word word_list: the list of words which are in our movie space docs: a list of new docs we're using to update our word co-occurence Returns ======= word_co, word_count_dict : new word co-occurence dict/mat and new word count dictionary """ # Get English stop words stop_words = stopwords.words('english') # Make init dict word_co = defaultdict(list) word_count_dict = defaultdict(int) word_co.update(word_co_old) word_count_dict.update(word_count_dict_old) # Find co occurences in context data, based on document (content) for doc in docs: # Double loop to count word co-occurences tkns = self.tokenizer.tokenize(punct_strip(doc)) for i in range(len(tkns)): if tkns[i] not in stop_words: word_count_dict[tkns[i]] += 1 for j in range(len(tkns)): if not (j == i) and (tkns[j] in word_list): word_co[tkns[i]] = update_word_counts(word_co[tkns[i]], tkns[j]) return word_co, word_count_dict def query_vectorize(self, q, sw=False): # Remove punctuation, lowercase, and encode to utf query = punct_strip(q.lower().encode("utf-8").decode("utf-8")) # Tokenize query and check query stopword cutoff query_words = self.tokenizer.tokenize(query) # Remove stop words if necessary stop_words = stopwords.words('english') # Get English stop words if (sw): new_query = [] for x in query_words: if x not in stop_words: new_query.append(x) query_words = new_query # Make query tfidf query_tfidf = defaultdict(int) for word in query_words: query_tfidf[word] += 1 for word in query_tfidf: if word in self.idf: query_tfidf[word] *= self.idf[word] else: query_tfidf[word] = 0 # Find query norm query_norm = 0 for word in query_tfidf: query_norm += math.pow(query_tfidf[word], 2) query_norm = math.sqrt(query_norm) return query_tfidf, query_norm def pseudo_rocchio(self, query_tfidf, query_norm, relevant, sw=False, a=.3, b=.4, clip=True): """ Arguments: query: a string representing the name of the movie being queried for relevant: a list of int representing the indices of relevant movies for query irrelevant: a list of strings representing the names of irrelevant movies for query a,b: floats, corresponding to the weighting of the original query, relevant queriesrespectively. clip: boolean, whether or not to clip all returned negative values to 0 Returns: q_mod: a dict representing the modified query vector. this vector should have no negatve weights in it! """ relevant_id = [] for s, i in relevant: relevant_id.append(i) if query_norm == 0: return self.find_random() # Calculate alpha*query_vec query_vec = query_tfidf for word in query_vec: query_vec[word] /= query_norm query_vec[word] *= a # Get words in relevant docs relevant_words = [] relevant_context = [] for i in relevant_id: relevant_context.append(self.context[i]) for context in relevant_context: context_tkns = self.tokenizer.tokenize(context) for tkn in context_tkns: if tkn not in relevant_words: relevant_words.append(tkn) # Collect relevant doc vector sums relevant_docs = defaultdict(int) for word in relevant_words: if word in self.inverted_index: for quote_id, tf in self.inverted_index[word]: if quote_id in relevant_id: relevant_docs[word] += (tf / self.norms[quote_id]) # Calculate beta term beta_term = b * (1.0 / len(relevant)) for key in relevant_docs: relevant_docs[key] *= beta_term # Sum query and relevant q_mod = {k: query_vec.get(k, 0) + relevant_docs.get(k, 0.0) for k in set(query_vec) | set(relevant_docs)} # negative checks for terms, if clip if (clip): for key in q_mod: if q_mod[key] < 0: q_mod[key] = 0 return q_mod else: return q_mod def find_random(self): r = random.randint(0, len(self.quotes)) return [[self.quotes[r], self.movies[r], self.context[r]]] def find_similar(self, query): query_words = self.tokenizer.tokenize(query) query_tfidf = defaultdict(int) for word in query_words: query_tfidf[word] += 1 for word in query_tfidf: if word in self.idf: query_tfidf[word] *= self.idf[word] else: query_tfidf[word] = 0 query_norm = 0 for word in query_tfidf: query_norm += math.pow(query_tfidf[word], 2) query_norm = math.sqrt(query_norm) if query_norm == 0: return self.find_random() scores = [0 for _ in self.quotes] for word in query_tfidf: if word in self.inverted_index: for quote_id, tf in self.inverted_index[word]: scores[quote_id] += query_tfidf[word] * tf * self.idf[word] results = [] for i, s in enumerate(scores): if self.norms[i] != 0: results.append((s / (self.norms[i] * query_norm), i)) top_res_num = 5 results.sort(reverse=True) return [[self.quotes[i], self.movies[i], self.context[i]] for _, i in results[:top_res_num]] def find_final(self, q, rocchio=True, pseudo_rocchio_num=5, sw=False, pmi_num=8, ml=False): """ Arguments: q: a string representing the query rocchio: a boolean representing whether or not to use pseudo relevance feedback with Rocchio psudo_rocchio_num: and int representing the number of top documents to consider relevant for rocchio sw: a boolean on whether or not to include stop words. pmi_num: an int representing the number of items to add to the query to expand it with PMI. Returns: result_quotes: a list of the top x results """ # Vectorize query query_tfidf, query_norm = self.query_vectorize(q, sw) if query_norm == 0: r = random.randint(0, len(self.quotes)) return [[self.quotes[r], self.movies[r], self.context[r]]] # Expand query using PMI # http://www.jofcis.com/publishedpapers/2011_7_1_17_24.pdf pmi_expansion = defaultdict(float) pmi_norm = 1 for word in query_tfidf: # Sum PMI lists if word in self.pmi_dict.keys(): pmi_list = self.pmi_dict[word][:pmi_num] pmi_score_list = [] for word, score in pmi_list: pmi_expansion[word] += score pmi_score_list.append(score) temp_norm = 0 for s in pmi_score_list: temp_norm += math.pow(s, 2) temp_norm = math.sqrt(query_norm) pmi_norm *= temp_norm query_tfidf.update(pmi_expansion) query_norm = query_norm * 2 * pmi_num * pmi_norm # Find query norm query_norm = 0 for word in query_tfidf: query_norm += math.pow(query_tfidf[word], 2) query_norm = math.sqrt(query_norm) # Get scores scores = [0 for _ in self.quotes] for word in query_tfidf: if word in self.inverted_index: for quote_id, tf in self.inverted_index[word]: scores[quote_id] += query_tfidf[word] * tf * self.idf[word] results = [] for i, s in enumerate(scores): if self.norms[i] != 0: results.append((s / (self.norms[i] * query_norm), i)) # Weight scores with year and rating for i in range(len(results)): score = results[i][0] index = results[i][1] year = self.year_rating_dict[self.movies[i]][0] rating = self.year_rating_dict[self.movies[i]][1] results[i] = (year_rating_weight(float(year), float(rating), score), index) # sort results results.sort(reverse=True) if rocchio: # Do pseudo-relevance feedback with Rocchio mod_query = self.pseudo_rocchio(query_tfidf, query_norm, results[:pseudo_rocchio_num], sw) mod_query_norm = 0 for word in mod_query: mod_query_norm += math.pow(mod_query[word], 2) mod_query_norm = math.sqrt(mod_query_norm) # Re-find scores and reweight with year and rating scores = [0 for _ in self.quotes] for word in mod_query: if word in self.inverted_index: for quote_id, tf in self.inverted_index[word]: scores[quote_id] += mod_query[word] * tf * self.idf[word] results = [] for i, s in enumerate(scores): if self.norms[i] != 0: results.append((s / (self.norms[i] * mod_query_norm), i)) d_score_updates = {} if ml is True: d_score_updates = self.find_ml(q) # Weight scores with year and rating for i in range(len(results)): score = results[i][0] index = results[i][1] year = self.year_rating_dict[self.movies[i]][0] rating = self.year_rating_dict[self.movies[i]][1] results[i] = (year_rating_weight(float(year), float(rating), score), index) if ml is True and index in d_score_updates: results[i] = (results[i][0]*0.9 + d_score_updates[index], results[i][1]) # Sort and return results top_res_num = 5 results.sort(reverse=True) used_quotes = [] return_res = [] counter = 0 while len(return_res) <= top_res_num: # Avoid duplicate quotes score, i = results[counter] if self.quotes[i] not in used_quotes: used_quotes.append(self.quotes[i]) return_res.append((score, i)) else: counter += 1 result_quotes = [[self.quotes[i], self.movies[i], self.context[i]] for _, i in return_res[:top_res_num]] return result_quotes def sentiment_analysis(self, td): with open('jsons/all_words_neg.pickle', 'rb') as f: all_words_neg = pickle.load(f) with open('jsons/training_docs.pickle', 'rb') as f: training_docs = pickle.load(f) genres = ['action', 'crime', 'comedy', 'drama'] testing_docs = [(td, genre) for genre in genres] all_words_neg = self.sentim_analyzer.all_words([mark_negation(doc) for doc in training_docs]) unigram_feats = self.sentim_analyzer.unigram_word_feats(all_words_neg, min_freq=4) all_words_neg = self.sentim_analyzer.all_words([mark_negation(doc) for doc in training_docs]) self.sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) training_set = self.sentim_analyzer.apply_features(training_docs) test_set = self.sentim_analyzer.apply_features(testing_docs) trainer = NaiveBayesClassifier.train classifier = self.sentim_analyzer.train(trainer, training_set) # f = open('my_classifier_test.pickle', 'rb') # classifier = pickle.load(f) # f.close() # classifier = nltk.data.load("my_classifier.pickle") genre_accuracy = [] for key, value in sorted(self.sentim_analyzer.evaluate(test_set).items()): # print('{0}: {1}'.format(key, value)) if key == 'Precision [action]': genre_accuracy.append(('action', value)) if key == 'Precision [comedy]': genre_accuracy.append(('comedy', value)) if key == 'Precision [drama]': genre_accuracy.append(('drama', value)) if key == 'Precision [crime]': genre_accuracy.append(('crime', value)) return genre_accuracy # Takes in a query # Outputs a dictionary of movie indices movies to weights where weight is to be added to all quote scores of movies def find_ml(self, td): f_tokenizer = TreebankWordTokenizer() query_words = f_tokenizer.tokenize(td) genres = self.sentiment_analysis(query_words) weighted_genres = [] genre_weights = {} for x in genres: if x[1] is not None: weighted_genres.append(x[0]) genre_weights[x[0]] = x[1] d_score_updates = {} for movie in self.movies: g = self.genre_dict[movie][0] total_genre_score = 0 if u'Comedy' in g and 'comedy' in weighted_genres: total_genre_score += genre_weights['comedy'] if u'Action' in g and 'action' in weighted_genres: total_genre_score += genre_weights['action'] if u'Crime' in g and 'crime' in weighted_genres: total_genre_score += genre_weights['crime'] if u'Drama' in g and 'drana' in weighted_genres: total_genre_score += genre_weights['drama'] d_score_updates[self.movies.index(movie)] = total_genre_score * .1 return d_score_updates
def demo_subjectivity(trainer, save_analyzer=False, n_instances=None, output=None): """ Train and test a classifier on instances of the Subjective Dataset by Pang and Lee. The dataset is made of 5000 subjective and 5000 objective sentences. All tokens (words and punctuation marks) are separated by a whitespace, so we use the basic WhitespaceTokenizer to parse the data. :param trainer: `train` method of a classifier. :param save_analyzer: if `True`, store the SentimentAnalyzer in a pickle file. :param n_instances: the number of total sentences that have to be used for training and testing. Sentences will be equally split between positive and negative. :param output: the output file where results have to be reported. """ from nltk.sentiment import SentimentAnalyzer from nltk.corpus import subjectivity if n_instances is not None: n_instances = int(n_instances / 2) subj_docs = [ (sent, "subj") for sent in subjectivity.sents(categories="subj")[:n_instances] ] obj_docs = [(sent, "obj") for sent in subjectivity.sents(categories="obj")[:n_instances]] # We separately split subjective and objective instances to keep a balanced # uniform class distribution in both train and test sets. train_subj_docs, test_subj_docs = split_train_test(subj_docs) train_obj_docs, test_obj_docs = split_train_test(obj_docs) training_docs = train_subj_docs + train_obj_docs testing_docs = test_subj_docs + test_obj_docs sentim_analyzer = SentimentAnalyzer() all_words_neg = sentim_analyzer.all_words( [mark_negation(doc) for doc in training_docs]) # Add simple unigram word features handling negation unigram_feats = sentim_analyzer.unigram_word_feats(all_words_neg, min_freq=4) sentim_analyzer.add_feat_extractor(extract_unigram_feats, unigrams=unigram_feats) # Apply features to obtain a feature-value representation of our datasets training_set = sentim_analyzer.apply_features(training_docs) test_set = sentim_analyzer.apply_features(testing_docs) classifier = sentim_analyzer.train(trainer, training_set) try: classifier.show_most_informative_features() except AttributeError: print( "Your classifier does not provide a show_most_informative_features() method." ) results = sentim_analyzer.evaluate(test_set) if save_analyzer == True: save_file(sentim_analyzer, "sa_subjectivity.pickle") if output: extr = [f.__name__ for f in sentim_analyzer.feat_extractors] output_markdown( output, Dataset="subjectivity", Classifier=type(classifier).__name__, Tokenizer="WhitespaceTokenizer", Feats=extr, Instances=n_instances, Results=results, ) return sentim_analyzer