Exploiting Contextual Similarity in Sentiment Analysis Using Word Embedding and Recurrent Neural Network

Abstract

Rapid growth of Internet users across the world and technological advancement in smart devices in last decade has brought many traditional public services on the Internet, leading them to avail these services sitting at home in their comfort zone and providing their feedback on social media or service providers web sites. Now people express their opinion freely, conveniently and frequently, resulting in a huge volume of opinion data on the Internet. Today’s era is quite competitive and, in this competition, if any organization want to stay relevant then they will have to pay enough attention on the users’ opinion and that’s what companies are doing by leveraging artificial intelligence and natural language processing. However, these opinion data are only available in unstructured format with lot of hidden information. Sentiment Analysis is one of the challenging tasks in NLP where with the help of machine learning algorithms and language modelling these opinion data are collected, processed and exploited to provide better services to the users. In this paper we have presented machine learning approach for exploiting contextual information in sentences. There are many machines learning based probabilistic approach to classify users’ reviews in positive and negative classes as per the polarity of the terms in text data, but most of them treats opinion words independently while in reality in any language words possess strong relationships with other words, resulting in low precision of the NLP model for sentiment analysis. Feature representation for the collected text data and how machine learning model learns from text are crucial to build a good model. With distributed word vector representation, this missing contextual information of the neighbouring words can be exploited in predicting the polarity of the reviews. Polarity of the sentence or paragraph cannot be seen as a count of individual terms on the basis of their independent positive or negative polarity but also it takes into consideration of those neighbouring words in the sentences that has major contribution in the overall sentiment of the review document. We used neural network to perform sentiment analysis on processed input user review data in a way that well trained model should be able to predict user sentiment with high accuracy .We experimented with the tweeter dataset (Go, Bhayani, and Huang n.d.) publicly available for NLP research. Experimental result shows that neural network model outperform other lexicon-based approaches in terms of accuracy when supplied with the contextual information.