An Evaluation of Tweet Sentiment Classification Methods

In this paper, we present the result of our research in predicting sentiment from Twitter data derived from a Kaggle competition. Our goal was to determine the efficacy of different supervised classification methods to predict Twitter sentiment to be Positive, Neutral or Negative. We evaluated four different classification statistical models: 1. Logistic Regression (LR), 2. Linear Support Vector Machine (LSV), 3. Multinomial Naïve Bayesian (NB), and 4. Random Forest (RF). We also evaluated two different tokenization methods 1. Document Term Matrix (DTM) and 2. Term Frequency-Inverse Document Frequency (TF-IDF). We combined this with three extraction methods 1. Original Tweet Text, 2. Rapid Automatic Keyword Extraction (RAKE) and 3. Hand curated Selected Text. Furthermore, various Neural Networks were applied to the Tweet Text and BERT extracted data that reduced the original 1000 features to be 768 that were applied to different models. Our experiment shows RF and LR gives the best results and there is little difference between DTM and TF-IDF. Fully connected neural network (FCNN) performed the best for the BERT extracted data with a test score of 0.75.