A Deep Learning Architecture with Word Embeddings to Classify Sentiment in Twitter

Abstract

Social Media Networks are one of the main platforms to express our feelings. The emotions we put in text tell a lot about our behavior towards any topic. Therefore, the analysis of text is a need for detecting one’s emotions in many fields. This paper introduces a deep learning model that classify sentiments from tweets using different types of word embeddings. The main component of our model is the Convolutional Neural Network (CNN) and the main used features are word embeddings. Trials are made on randomly initialized word embeddings and pretrained ones. The used pre-trained word embeddings are of different variants such as Word2Vec, Glove and fastText models. The model consists of three CNN streams that are concatenated and followed by a fully-connected layer. Each stream contains only one convolutional layer and one max-pooling layer. The model works on detecting positive and negative emotions from Stanford Twitter Sentiment (STS) dataset. The accuracy achieved is 78.5% when using the randomly initialized word embeddings and achieved a maximum accuracy 84.9% using Word2Vec word embeddings. The model not only proves that randomly initialized word embedding can achieve good accuracy, it also showing the power of the pretrained word embeddings that helps to achieve a higher competitive accuracy in sentiment classification.