Yelp2018SentimentAnalysis

Insights:

Strippping punctuation seems to increase performance by 1% -> Better word tokenizing (Try Punkt tokenizer trained on text from nltk instead of default)

Class-balancing nukes the accuracy by 5%.

Ranking:

1. 65.8%~~

70 words, no punctuation, no stopwords, FB vectors, CNN+LSTM

Epoch 200+ - 27s - loss: 0.4310 - percentageCorrect: 65.9728 - val_loss: 0.4347 - val_percentageCorrect: 65.7453

embedding_1 (Embedding) (None, 70, 300) 57183900

dropout_1 (Dropout) (None, 70, 300) 0

conv1d_1 (Conv1D) (None, 66, 32) 48032

max_pooling1d_1 (MaxPooling1 (None, 13, 32) 0

lstm_1 (LSTM) (None, 70) 28840

dense_1 (Dense) (None, 200) 14200

dropout_2 (Dropout) (None, 200) 0

dense_2 (Dense) (None, 200) 40200

Total params: 57,315,373 Trainable params: 57,315,373 Non-trainable params: 0

Predicting: Stars, Funny, Useful, Cool

Ranking:

1. 66.2% accuracy

80 words, no punctuation, no stopwords, FB vectors, CNN

Epoch 123/200

• 12s - loss: 0.6433 - percentageCorrect: 66.8893 - val_loss: 0.6942 - val_percentageCorrect: 66.2160

model = Sequential() model.add(embedding_layer) model.add(Conv1D(filters=32, kernel_size=10, padding='same', activation='relu')) model.add(MaxPooling1D(pool_size=2)) SpatialDropout1D(0.5) model.add(Conv1D(filters=16, kernel_size=10, padding='same', activation='relu')) model.add(MaxPooling1D(pool_size=2)) model.add(Conv1D(filters=8, kernel_size=10, padding='same', activation='relu')) model.add(MaxPooling1D(pool_size=2)) SpatialDropout1D(0.5) model.add(Conv1D(filters=8, kernel_size=10, padding='same', activation='relu')) model.add(MaxPooling1D(pool_size=2)) SpatialDropout1D(0.5) model.add(Flatten()) model.add(Dense(512, activation='relu')) model.add(Dropout(0.8)) model.add(Dense(512, activation='relu')) model.add(Dropout(0.8)) model.add(Dense(256, activation='relu')) model.add(Dropout(0.8)) model.add(Dense(128, activation='relu')) model.add(Dense(4)) model.compile(loss='mean_absolute_error', optimizer='adam', metrics=[percentageCorrect])

2. 65.34% Accuracy

80 words, puntuaction, no stopwords FB vectors, CNN

Epoch 87/200

• 11s - loss: 0.6843 - percentageCorrect: 65.4592 - val_loss: 0.6861 - val_percentageCorrect: 65.3440

model = Sequential() model.add(embedding_layer) model.add(Conv1D(filters=32, kernel_size=10, padding='same', activation='relu')) model.add(MaxPooling1D(pool_size=2)) SpatialDropout1D(0.5) model.add(Conv1D(filters=16, kernel_size=10, padding='same', activation='relu')) model.add(MaxPooling1D(pool_size=2)) model.add(Conv1D(filters=8, kernel_size=10, padding='same', activation='relu')) model.add(MaxPooling1D(pool_size=2)) SpatialDropout1D(0.5) model.add(Conv1D(filters=8, kernel_size=10, padding='same', activation='relu')) model.add(MaxPooling1D(pool_size=2)) SpatialDropout1D(0.5) model.add(Flatten()) model.add(Dense(512, activation='relu')) model.add(Dropout(0.8)) model.add(Dense(512, activation='relu')) model.add(Dropout(0.8)) model.add(Dense(256, activation='relu')) model.add(Dropout(0.8)) model.add(Dense(128, activation='relu')) model.add(Dense(4)) model.compile(loss='mean_absolute_error', optimizer='adam', metrics=[percentageCorrect])