LSTM based Deep RNN Architecture for Election Sentiment Analysis from Bengali Newspaper

Abstract

This work presents the sentiment analysis of the political news articles in Bengali. This domain is quite complicated, because the political news is subjective and depends on several socio economic factors and largely over the geographical areas, communities, etc. The work is conducted in two steps the pre-processing and the classifications. In the pre-processing phase, the data is prepared. It is an automated process, which collects data from the web and identifies the relevant news article. A special type of Recurrent Neural Network (RNN) based deep learning approach called Long Short Term Memory (LSTM) is used for election sentiment classification and the result is compared with other supervised classifiers like Naive Bayes, SVM, and Decision Tree. Different types of word embedding models are compared for election sentiment analysis. Results demonstrate that LSTM based deep RNN architecture with context encoder provides 85% of accuracy which dominates others. The result shows that it is close to the existing state-of-the-art and it also gives very attractive insights.