Comparaison between the implementation of emotion detection from Twitter Tweets using SVM and LSTM

Abstract

This report compares two di?erent Machine Learning (ML) methods used to classify five types of emotions from a twitter tweets dataset.

The first approach is a classical method in Natural Language Processing (NLP), Support Vector Machine (SVM); The text data is cleaned, tokenized, and stemmed to derive fea-ture vectors using two di?erent feature extraction methods, namely Bag of Words (BoW) and Term Frequency-Inverse Document Frequency (TF-IDF). The resulting feature ma-trix is fed to a non-linear SVM classifier. Conversely, the second approach is more recent; this method uses word embedding and Long Short Term Memory (LSTM) neural network. First, we convert words of similar meaning into similar feature vectors. Then, the result-ing features are fed sequentially into the LSTM.

Although it has been proved in the past that the SVM is the most robust model in classifica-tion problems, it is not the case for text classification. LSTM showed a better performance compared to the SVM; between 85% and 87% for LSTM and around 82% for the SVM.