Text Classification for Sentiment Prediction of Social Media Dataset using Multichannel Convolution Neural Network

Abstract

Text classification of social media documents requires a specified model for a particular dataset. We investigated the public opinion and sentiment from Indonesian twitter toward the employment problem. To predict the sentiment about this topic, we collected a total of 75,126 tweets which correlate with employment conversation for the training and testing process. We analyzed the collected tweets by applying convolution neural networks (CNN) model which expanded from the standard one-dimensional deep learning model to a multichannel model. Applying multiple parallel CNN processes, the documents were processed using different CNN topologies by varying word embedding dimension, kernel size, number of hidden layers or nodes, and also different n-grams. Experiment results with three-dimension channel have 0.0172 loss error and 0.9917 accuracy. CNN model with multichannel configuration provides maximum accuracy and small loss error in training and testing process of model development.