Sentiment Analysis Using Out of Core Learning

Text sentiment detection for a particular language other than English is one of the challenging tasks presently. The reasons are; it needs a large dataset, language has no specific structure, one word has a different meaning, and it is hard for even human to understand the connotation of particular words. There exists several proposed architecture for detecting emotions in the Bengali language using machine learning and deep learning approaches, but they are not accurate enough to predict the perfect emotion of the sentence. And there is still no standalone architecture is available that can extract the sentiments hidden inside of a sentence in different languages. In this paper, we are proposing an abstract model that can enable sentiment analysis without any restriction of using a fixed language somewhat applicable to any language. With the use of natural language processing, we have extracted the features, and these features are then fed to different machine learning models for classification. As our main concern was to build up a general model, this model is confined to binary classification, i.e., positive and negative. Apart from this, In our system architecture, we have implemented stochastic gradient descent for optimization. So our model can be called out of core learning model where the model can be updated when new user data is inserted without training the whole model. For the evaluation of the performance of our model, we have trained the estimators against Bangla translated IMDB review dataset and calculated different evaluation metrics for our estimators. The dataset is translated into Bangla using google translator.