COMPOSITE WORD EMBEDDINGS BASED IMPROVED SENTIMENT ANALYSIS USING DEEP LEARNING

Abstract

Due to the explosive growth and popularity of Social Media sites, massive amounts of raw data are available that can be used for opinion mining and other pattern identification tasks. Identifying and summarizing the opinion or sentiment regarding any particular topic can be used to provide insights and can be taken as feedback to improve or address concerns regarding that topic. Most of the work done so far has focused on run of the mill, well-defined techniques like SVM, LDA and other machine learning techniques to classify a particular sample into one of several classes that indicates the sentiment inherently expressed in that sample. However, these techniques peak out, in terms of accuracy at a certain limit. Additional improvements to this accuracy has been reported when using deep neural networks. Segmenting dataset into multiple classes based on number or entities and the use of pre-trained word embeddings along with CNN has given the best results so far. The aim of the present work is to improve existing state of the art techniques based on word embeddings by leveraging the complementary nature of different kinds of word embeddings for sentiment classification. The performance of various models is evaluated on “Fine Food”, “IMDB Movie review”, “Yelp reviews” and “Twitter” datasets. The results obtained show that the proposed approach gives promising results