Multi-layers Convolutional Neural Network for Twitter Sentiment Ordinal Scale Classification

Abstract

Twitter sentiment analysis according to five points scales has attracted research interest due to its potential use in commercial and public social media application. A multi-point scale classification is a popular way used by many companies to evaluate the sentiment of product reviews (e.g. Alibaba, Amazon and eBay). Most of the classification approaches addressed this problem using traditional classification algorithm that requires expert knowledge to select the best features. Even though deep learning has been utilized, most of them employed a simple structure that not enough to capture the important features. In this paper, a complex structure of convolutional neural network (CNN) is proposed to classify the tweet into five-point scale and obtain a more several tweet representation. After a series of experiments with CNN including different hyperparameters and pooling strategies (Max and Average), we found that the best structure for our model is three convolutional layers, each one followed by average pooling layer. The proposed multi-layers convolutional neural network (MLCNN) model achieve the lowest Macro average mean absolute error (MAE^M) and outperforms the state-of-the-art approach on tweet 2016 dataset for Ordinal classification. Experimental results show the ability of average pooling to preserve significant features that provide more expressiveness to ordinal scale.