Identifying Sentiment of Malayalam Tweets Using Deep Learning

Abstract

The current chapter focus on providing a comparative study for identifying sentiment of Malayalam tweets using deep learning methods such as convolutional neural net (CNN), long short-term memory units (LSTM). The baseline methods used to compare are support vector machines (SVM), regularized least square classification with random kitchen sink mapping (RKS-RLSC). Malayalam is a low resource language spoken in Kerala state, India. Due to the unavailability of data, tweets were collected and labeled manually based on its polarity as neutral, negative and positive. RKS mapping is a well explored approach in which data are nonlinearly mapped to higher dimension where linear classifier can be used. The evaluation measure chosen for the experiments are F1-score, recall, accuracy and precision. The experiments also provide a comparison with classical methods such as logistic regression (LR), adaboost (Ab), random forest (RF), decision tree (DT), k-nearest neighbor (KNN) on the basis of accuracy as the measure. For the experiments using CNN and LSTM, we report the effectiveness of activation functions such as rectified linear units (ReLU), exponential linear units (ELU) and scaled exponential linear units (SELU) for the sentiment identification of Malayalam tweets over SVM and RKS-RLSC.