Collaborated with two team members to build models based on Long Short-Term Memory (LSTM) and Bidirectional Encoder Representations from Transformers (BERT) to predict the overall sentiment (positive, negative, or neural) in tweet dataset, and compare the results. I focused on the LSTM model implementation. Used a training dataset with 27471 tweets and testing dataset with 3534 tweets from Kaggle. Achieved training accuracy up to 88%.
Skills: LSTM, BERT, Keras, Pandas, sci-kit learn, Google Colab
a. Remove all rows containing null values.
b. Remove all punctuations and special symbols except those that are meaningful for sentiment
extraction (For example: **** in the sentence of "Son of ****" helps us predict the sentiment as
negative.). As a result, we can obtain more precise training result.
c. Convert all characters to lower case so the same letter with different cases will be seen as the same.
d. Tokenize every tweet into separate word using the tokenizer in the deep learning library Keras.
e. Convert distinct words and labels into numeric values (for example: ‘son’, ‘you’, ‘whether’ are
converted to number 1, 2,3). Note: we combine the distinct tokenized words in training set and test set before converting them to numbers in order to make sure these 2 data sets have same dimensions of tokenized text.
