sunny-side-up

Lab41's foray into Sentiment Analysis with Deep Learning. In addition to checking out the source code, visit the Wiki for Learning Resources and possible Conferences to attend.

Try them, try them, and you may! Try them and you may, I say.

Table of Contents

• Blog Overviews
• Docker Environments
• Binary Classification with Word Vectors
  • Word Vector Models
  • Training and Testing Data
• Binary Classification via Deep Learning
  • CNN
  • LSTM

Blog Overviews

Can Word Vectors Help Predict Whether Your Chinese Tweet Gets Censored?	March 2016
One More Reason Not To Be Scared of Deep Learning	March 2016
Some Tips for Debugging in Deep Learning	January 2016
Faster On-Ramp to Deep Learning With Jupyter-driven Docker Containers	November 2015
A Tour of Sentiment Analysis Techniques: Getting a Baseline for Sunny Side Up	November 2015
Learning About Deep Learning!	September 2015

Docker Environments

• lab41/itorch-[cpu|cuda]: iTorch IPython kernel for Torch scientific computing GPU framework
• lab41/keras-[cpu|cuda|cuda-jupyter]: Keras neural network library (CPU or GPU backend from command line or within Jupyter notebook)
• lab41/neon-[cuda|cuda7.5]: neon Deep Learning framework (with CUDA backend) by Nervana
• lab41/pylearn2: pylearn2 machine learning research library
• lab41/sentiment-ml: build word vectors (Word2Vec from gensim; GloVe from glove-python), tokenize Chinese text (jieba and pypinyin), and tokenize Arabic text (NLTK and Stanford Parser)
• lab41/mechanical-turk: convert CSV of Arabic tweets to individual PNG images for each Tweet (to avoid machine-translation of text) and auto-submit/score Arabic sentiment survey via AWS Mechanical Turk

Binary Classification with Word Vectors

Execution

python -m benchmarks/baseline_classifiers

Word Vector Models

model	filename	filesize	vocabulary	details
Sentiment140	sentiment140_800000.bin	153M	83,586	gensim Word2Vec(size=200, window=5, min_count=10)
Open Weiboscope	openweibo_fullset_hanzi_CLEAN_vocab31357747.bin	56G	31,357,746	jieba-tokenized Hanzi Word2Vec(size=200, window=5, min_count=1)
Open Weiboscope	openweibo_fullset_min10_hanzi_vocab2548911.bin	4.6G	2,548,911	jieba-tokenized Hanzi Word2Vec(size=200, window=5, min_count=10)
Arabic Tweets	arabic_tweets_min10vocab_vocab1520226.bin	1.2G	1,520,226	Stanford Parser-tokenized Word2Vec(size=200, window=5, min_count=10)
Arabic Tweets	arabic_tweets_NLTK_min10vocab_vocab981429.bin	759M	981,429	NLTK-tokenized Word2Vec(size=200, window=5, min_count=10)

Training and Testing Data

train/test set	filename	filesize	details
Sentiment140	sentiment140_800000_samples_[test/train].bin	183M	80/20 split of 1.6M emoticon-labeled Tweets
Open Weiboscope	openweibo_hanzi_censored_27622_samples_[test/train].bin	25M	80/20 split of 55,244 censored posts
Open Weiboscope	openweibo_800000_min1vocab_samples_[test/train].bin	564M	80/20 split of 1.6M deleted posts
Arabic Twitter	arabic_twitter_1067972_samples_[test/train].bin	912M	80/20 split of 2,135,944 emoticon-and-emoji labeled Tweets

Binary Classification via Deep Learning

CNN (Convolutional Neural Network)

Character-by-character processing From Zhang and LeCun's Text Understanding From Scratch:

#Set Parameters for final fully connected layers
fully_connected = [1024,1024,1]

model = Sequential()

#Input = #alphabet x 1014
model.add(Convolution2D(256,67,7,input_shape=(1,67,1014)))
model.add(MaxPooling2D(pool_size=(1,3)))

#Input = 336 x 256
model.add(Convolution2D(256,1,7))
model.add(MaxPooling2D(pool_size=(1,3)))

#Input = 110 x 256
model.add(Convolution2D(256,1,3))

#Input = 108 x 256
model.add(Convolution2D(256,1,3))

#Input = 106 x 256
model.add(Convolution2D(256,1,3))

#Input = 104 X 256
model.add(Convolution2D(256,1,3))
model.add(MaxPooling2D(pool_size=(1,3)))

model.add(Flatten())

#Fully Connected Layers

#Input is 8704 Output is 1024
model.add(Dense(fully_connected[0]))
model.add(Dropout(0.5))
model.add(Activation('relu'))

#Input is 1024 Output is 1024
model.add(Dense(fully_connected[1]))
model.add(Dropout(0.5))
model.add(Activation('relu'))

#Input is 1024 Output is 1
model.add(Dense(fully_connected[2]))
model.add(Activation('sigmoid'))

#Stochastic gradient parameters as set by paper
sgd = SGD(lr=0.01, decay=1e-5, momentum=0.9, nesterov=True)
model.compile(loss='binary_crossentropy', optimizer=sgd, class_mode="binary")

LSTM (Long Short Term Memory)

# initialize the neural net and reshape the data
model = Sequential()
model.add(Embedding(max_features, embedding_size)) # embed into dense 3D float tensor (samples, maxlen, embedding_size)
model.add(Reshape(1, maxlen, embedding_size)) # reshape into 4D tensor (samples, 1, maxlen, embedding_size)

# convolution stack
model.add(Convolution2D(nb_feature_maps, nb_classes, filter_size_row, filter_size_col, border_mode='full')) # reshaped to 32 x maxlen x 256 (32 x 100 x 256)
model.add(Activation('relu'))

# convolution stack with regularization
model.add(Convolution2D(nb_feature_maps, nb_feature_maps, filter_size_row, filter_size_col, border_mode='full')) # reshaped to 32 x maxlen x 256 (32 x 100 x 256)
model.add(Activation('relu'))
model.add(MaxPooling2D(poolsize=(2, 2))) # reshaped to 32 x maxlen/2 x 256/2 (32 x 50 x 128)
model.add(Dropout(0.25))

# convolution stack with regularization
model.add(Convolution2D(nb_feature_maps, nb_feature_maps, filter_size_row, filter_size_col)) # reshaped to 32 x 50 x 128
model.add(Activation('relu'))
model.add(MaxPooling2D(poolsize=(2, 2))) # reshaped to 32 x maxlen/2/2 x 256/2/2 (32 x 25 x 64)
model.add(Dropout(0.25))

# fully-connected layer
model.add(Flatten())
model.add(Dense(nb_feature_maps * (maxlen/2/2) * (embedding_size/2/2), fully_connected_size))
model.add(Activation("relu"))
model.add(Dropout(0.50))

# output classifier
model.add(Dense(fully_connected_size, 1))
model.add(Activation("sigmoid"))

# try using different optimizers and different optimizer configs
model.compile(loss='binary_crossentropy', optimizer='adam', class_mode="binary")