def main(arguments):
# enable logging
logging.basicConfig(format='%(asctime)s : %(levelname)s : %(message)s',
filename=LOG_FNAME,
level=logging.INFO)
# parse optional filename arguments
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('-s',
'--sartic-tweets',
dest='sar_dir',
help="Directory of example sartic tweets",
default="../data/labeled_data/sarcastic/")
parser.add_argument('-p',
'--positive-tweets',
dest='pos_dir',
help="Directory of example positive tweets",
default="../data/labeled_data/positive/")
parser.add_argument('-n',
'--negative-tweets',
dest='neg_dir',
help="Directory of example negative tweets",
default="../data/labeled_data/negative/")
parser.add_argument(
'-c',
'--sample-count',
dest='sample_count',
help="Max number of samples of each class",
default="10000") # 10k default, ~300k max with current data
args = parser.parse_args(arguments)
# set random seed
np.random.seed(RAND_SEED)
# create tweets iterators
log_print("Creating tweet iterators...")
sar_tweets_iter = Tweets([args.sar_dir])
pos_tweets_iter = Tweets([args.pos_dir])
neg_tweets_iter = Tweets([args.neg_dir])
log_print()
# load tweets with gold labels filtered to lists and shuffle
log_print("Loading sarcastic tweets with gold labels filtered...")
sar_tweets = [Tweets.filter_tags(tweet) for tweet in sar_tweets_iter]
log_print("...loaded {} sarcastic tweets".format(len(sar_tweets)))
log_print("Loading non-sarcastic tweets...")
pos_tweets = [Tweets.filter_tags(tweet)
for tweet in pos_tweets_iter] # filter gold label hashtags
log_print("...loaded {} positive tweets...".format(len(pos_tweets)))
neg_tweets = [Tweets.filter_tags(tweet) for tweet in neg_tweets_iter]
log_print("...loaded {} negative tweets".format(len(neg_tweets)))
log_print(
"Selecting balanced sample sets of {} tweets per class...".format(
args.sample_count))
sample_count = int(args.sample_count)
sar_tweets = resample(sar_tweets,
n_samples=sample_count,
replace=False,
random_state=1)
pos_tweets = resample(pos_tweets,
n_samples=sample_count // 2,
replace=False,
random_state=2)
neg_tweets = resample(neg_tweets,
n_samples=sample_count // 2,
replace=False,
random_state=3)
non_tweets = pos_tweets + neg_tweets
log_print()
# shuffle tweets and split into training, dev, and test
log_print("Shuffle all tweets...")
sar_labels = [1 for _ in sar_tweets]
non_labels = [0 for _ in non_tweets]
tweets = np.append(sar_tweets, non_tweets)
labels = np.append(sar_labels, non_labels)
tweets, labels = shuffle(tweets, labels, random_state=4)
log_print()
# write to output file
log_print("write to files as training, dev, and test sets...")
output_gen = (n for n in zip(tweets, labels)
) # generator of (tweet, label) tuples
with open(OUTFNAME_FORMAT.format("test"), "w+") as f:
for tweet, label in itertools.islice(output_gen, sample_count // 10):
f.write("{}\t{}\n".format(label, ' '.join(tweet)))
with open(OUTFNAME_FORMAT.format("dev"), "w+") as f:
for tweet, label in itertools.islice(output_gen, sample_count // 10):
f.write("{}\t{}\n".format(label, ' '.join(tweet)))
with open(OUTFNAME_FORMAT.format("train"), "w+") as f:
for tweet, label in output_gen:
f.write("{}\t{}\n".format(label, ' '.join(tweet)))
log_print(
"...training, dev, and test sets written to files {}, {}, and {}".
format(OUTFNAME_FORMAT.format("train"), OUTFNAME_FORMAT.format("dev"),
OUTFNAME_FORMAT.format("test")))
def main(pos_dir, neg_dir, sar_dir, random_seed):
np.random.seed(random_seed)
# Create tweets iterators
update("Creating tweet iterators...")
pos_tweets_iter = Tweets([pos_dir])
neg_tweets_iter = Tweets([neg_dir])
sar_tweets_iter = Tweets([sar_dir])
update()
# Save situtations to lists and shuffle
update("Loading positive tweets...")
pos_tweets = [
' '.join(Tweets.filter_tags(tweet)) for tweet in pos_tweets_iter
]
pos_tweets = shuffle(pos_tweets)
update()
update("Loading negative tweets...")
neg_tweets = [
' '.join(Tweets.filter_tags(tweet)) for tweet in neg_tweets_iter
]
neg_tweets = shuffle(neg_tweets)
update()
update("Loading sarcastic tweets...")
sar_tweets = [
' '.join(Tweets.filter_tags(tweet)) for tweet in sar_tweets_iter
]
sar_tweets = shuffle(sar_tweets)
update()
# Save sarcasm data
update("Saving sarcasm data...")
count = len(sar_tweets)
print("len pos_tweets before take = {}".format(len(pos_tweets)))
non_sar_tweets = take(pos_tweets, count // 2) + take(
neg_tweets, count // 2)
print("len pos_tweets after take = {}".format(len(pos_tweets)))
sar_labels = [1 for _ in sar_tweets]
non_sar_labels = [0 for _ in non_sar_tweets]
sarcasm_data = np.append(sar_tweets, non_sar_tweets)
sarcasm_labels = np.append(sar_labels, non_sar_labels)
sarcasm_data, sarcasm_labels = shuffle(sarcasm_data, sarcasm_labels)
size = len(sarcasm_data)
train = slice(0, int(0.8 * size))
dev = slice(int(0.8 * size), int(0.9 * size))
test = slice(int(0.8 * size), size - 1)
sarcasm_dump = {
"train": (sarcasm_data[train], sarcasm_labels[train]),
"dev": (sarcasm_data[dev], sarcasm_labels[dev]),
"test": (sarcasm_data[test], sarcasm_labels[test])
}
pickle.dump(sarcasm_dump,
open(os.path.join(SPLIT_DATA_DIR, "sarcasm.pkl"), 'wb'))
update()
# Save sentiment data
update("Saving sentiment data...")
count = min(len(pos_tweets), len(neg_tweets))
pos_tweets = pos_tweets[:count]
neg_tweets = neg_tweets[:count]
pos_labels = [1 for _ in pos_tweets]
neg_labels = [0 for _ in neg_tweets]
sentiment_data = np.append(pos_tweets, neg_tweets)
sentiment_labels = np.append(pos_labels, neg_labels)
sentiment_data, sentiment_labels = shuffle(sentiment_data,
sentiment_labels)
size = len(sentiment_data)
train = slice(0, int(0.8 * size))
dev = slice(int(0.8 * size), int(0.9 * size))
test = slice(int(0.8 * size), size - 1)
sentiment_dump = {
"train": (sentiment_data[train], sentiment_labels[train]),
"dev": (sentiment_data[dev], sentiment_labels[dev]),
"test": (sentiment_data[test], sentiment_labels[test])
}
pickle.dump(sentiment_dump,
open(os.path.join(SPLIT_DATA_DIR, "sentiment.pkl"), 'wb'))
update()
def main(arguments):
# Parse optional filename arguments
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument('-p', '--positive-tweets', dest='pos_dir',
help="Directory of example positive tweets",
default="../data/labeled_data/positive/")
parser.add_argument('-n', '--negative-tweets', dest='neg_dir',
help="Directory of example negative tweets",
default="../data/labeled_data/negative/")
parser.add_argument('-c', '--sample-count', dest='sample_count',
help="Max number of samples of each sentiment",
default="800000")
args = parser.parse_args(arguments)
# Create Tweets Iterators
update("Creating tweet iterators...")
pos_tweets_iter = Tweets([args.pos_dir])
neg_tweets_iter = Tweets([args.neg_dir])
update()
# Save situtations to lists and shuffle
update("Loading positive tweets...")
pos_tweets = [' '.join(Tweets.filter_tags(tweet)) for tweet in pos_tweets_iter]
update()
update("Loading negative tweets...")
neg_tweets = [' '.join(Tweets.filter_tags(tweet)) for tweet in neg_tweets_iter]
update()
update("Selecting balanced sample sets...")
sample_count = int(args.sample_count)
pos_tweets = resample(pos_tweets, n_samples=sample_count,
replace=False, random_state=1)
neg_tweets = resample(neg_tweets, n_samples=sample_count,
replace=False, random_state=2)
update()
# Shuffle tweets and split into training, dev, and test
update("Shuffle tweets and split into training, dev, and test sets...")
pos_labels = [1 for _ in pos_tweets]
neg_labels = [0 for _ in neg_tweets]
tweets = np.append(pos_tweets, neg_tweets)
labels = np.append(pos_labels, neg_labels)
tweets, labels = shuffle(tweets, labels, random_state=2)
size = len(labels)
train = slice(0, int(0.8 * size))
dev = slice(int(0.8 * size), int(0.9 * size))
test = slice(int(0.8 * size), size - 1)
update()
print()
# Build Pipeline
print("Performing grid search...")
pipeline = Pipeline([('vect', CountVectorizer()),
#('tfidf', TfidfTransformer()),
('clf', MultinomialNB())])
parameters = { #TODO check which parameters actually effect use in sarcasm detection
'vect__tokenizer': [tokenizer],
'vect__stop_words': [None],
'vect__binary': [False],
'vect__ngram_range': [(1,5)],
#'tfidf__norm': [None, 'l1', 'l2'],
#'tfidf__use_idf': [True, False],
#'tfidf__smooth_idf': [True, False],
#'tfidf__sublinear_tf': [True, False],
'clf__alpha': [1.0], # check range, these are guesses
'clf__fit_prior': [False], # not sure what the distribution in sarcasm data is
}
clf_pipe = GridSearchCV(pipeline, parameters, n_jobs=-1, verbose=1)
print("pipeline:", [name for name, _ in pipeline.steps])
print("parameters:")
pprint(parameters)
t0 = time()
clf_pipe.fit(tweets[train], labels[train])
print("Done in %0.3fs" % (time() - t0))
print()
# Print grid search results
print("Best score: %0.3f" % clf_pipe.best_score_)
print("Best parameters set:")
best_parameters = clf_pipe.best_estimator_.get_params()
for param_name in sorted(parameters.keys()):
print("\t%s: %r" % (param_name, best_parameters[param_name]))
print()
# Evaluate classifier
vect = clf_pipe.best_estimator_.named_steps['vect']
#tfidf = clf_pipe.best_estimator_.named_steps['tfidf']
clf = clf_pipe.best_estimator_.named_steps['clf']
predicted = clf_pipe.predict(tweets[test])
print("Classifier Evaluation:")
print(metrics.classification_report(labels[test], predicted,
target_names=["-", "+"]))
# save classifier
pickle.dump(clf_pipe, open(MODEL_FNAME, 'wb'))
