International Conference on Bangla Speech and Language Processing(ICBSLP), 21-22 September, 2018 A Deep Recurrent Neural Network with BiLSTM model for Sentiment Classification Abdullah Aziz Sharfuddin Department of CSE Shahjalal University of Science and Technology Sylhet, Bangladesh abdullahshakkhor@gmail.com Md. Nafis Tihami Department of CSE Shahjalal University of Science and Technology Sylhet, Bangladesh nafistiham@student.sust.edu Md. Saiful Islam Department of CSE Shahjalal University of Science and Technology Sylhet, Bangladesh saiful-cse@sust.edu Abstract—In the field of sentiment classification, opinions or sentiments of the people are analyzed. Sentiment analysis systems are being applied in social platforms and in almost every business because the opinions or sentiments are the reflection of the beliefs, choices and activities of the people. With these systems it is possible to make decisions for businesses to political agendas. In recent times a huge number of people share their opinions across the Internet using Bengali. In this paper a new way of sentiment classification of Bengali text using Recurrent Neural Network(RNN) is presented. Using deep recurrent neural network with BiLSTM, the accuracy 85.67% is achieved. Index Terms—Bengali text; Deep learning; Sentiment Classifi- cation; RNN; LSTM; BiLSTM; Facebook; NLP I. I NTRODUCTION With the elevation in the communication technology i.e. world wide web, a huge number of people from all lineages across the world take part in social networks and express their emotions or opinions on a wide range of topics. Now it is a dire need to summarize the data created by people over the social networks and see the insights from them. Besides, in the field of NLP, it has become a topic of enormous interest. Because, it is needed to make smart recommending systems, anticipating the results of political elections, understanding the feedback of people on public events and movements. SA is a method of finding and classifying opinions ex- pressed in a piece of text basing on computation technologies, especially in order to find out whether the writer’s behavior towards a specific topic, product, etc. is positive, negative, or neutral. SA also refers to the administration of opinions, sentiments and subjective text [1]. It also gives the compre- hensive data associated to public views, as it goes through all the different kinds of tweets, reviews and comments. It is a verified mechanism for the prediction of a numerous momentous circumstances, for instance movie ratings at box office and public or regional elections [2]. Public opinions are used to value a certain matter, i.e. person, product or place and might be found at different websites like Amazon and Yelp. The sentiments can be specified into negative, positive or neutral and even more classes. SA can automatically find out the expressive direction of user reviews or opinions whether the users have a good or positive impression or a negative impression. [3] The usage of SA is broad and powerful. Its demand has grown due to the escalating need of extracting and inspecting hidden information from data coming through social medias. Different organizations around the world are using the ability to extract hidden data now-a-days. The change in sentiments can be connected to the change in stock market. The Obama administration used opinion mining in the 2012 presidential election to detect the public opinion before the announcement of policy. [4] Deep learning has shown great performance in SA. In this article, a way of SA using deep recurrent neural network with BiLSTM is presented. II. RELATED WORKS Sentiment analysis is not new for English language. A significant number of research works have been done within this scope. Arvind et al. [5] applied Skip-Gram model where as Paramveer et al. [6] applied CCA(Canonical Correlation Anal- ysis) for in-depth vectorization. Duyu Tang et al. [7] did their work on sentiment analysis of tweets. These contained not only the information but also the syntactic context.Bengali hasn’t yet progressed in this particular research area. Approaches that have been already made are dependent on machine learning mostly. Dipankar Das [8] used Parts of speech tagger for to tag emotion. He achieved 70% accuracy. K.M Azharul Hasan et al. [9] used contextual valency analysis in Bangla text for SA. In their approach, using parts of speech tagger, they calculated the three classes (positivity, negativity and neutrality).The percentages was summed up to get result. They got an accuracy of 75%. K.M Azharul Hasan et. al [10] used unsupervised learning algorithm and created phrase pattern. Shaika Chowdhury [11] used SVM and Maximum Entropy on Bengali micro blog posts. They compared these classifiers using the accuracy metric. M Al-Amin et. al [12] used word2vec model for their base of sentiment analysis in Bangla. They used the positions of vector representation of the Bengali words to create a positive or negative score for each word and determine the sentiment of a specific text. They achieved a 72.5% accuracy. Sentiment analysis on the rohingya issue was done by Hemayet et al. [13] Similar works were also done by Al-Amin et al. [14]. 978-1-5386-8207-4/18/$31.00 c 2018 IEEE