Open AccessUsense: A User Sentiment Analysis Model for Movie Reviews by Applying LSTM
Author(s) -
Undru Taran
Publication year - 2020
Publication title -
international journal for research in engineering application and management
Language(s) - English
Resource type - Journals
ISSN - 2454-9150
DOI - 10.35291/2454-9150.2020.0315
Subject(s) - sentiment analysis , computer science , support vector machine , sentence , mood , artificial intelligence , set (abstract data type) , machine learning , natural language processing , field (mathematics) , recurrent neural network , kernel (algebra) , training set , artificial neural network , psychology , mathematics , combinatorics , psychiatry , pure mathematics , programming language
Sentiment analysis is a field which deals with assessing the sentiments or emotions of the users on products and services. It takes user comments as input and applies natural language processing techniques to identify the mood of the user. Usually a sentiment is deemed to be positive, negative or neutral depending upon the mood that he expresses in the comments or feedbacks. It is largely used by businesses to improve products and services and also to present its customers with a set of products and services based on their likes and dislikes. State-of-the-art indicates many techniques have been applied in past such as, linear regression and SVM models. Recurrent Neural Networks (RNNs) have improved the way in which sentiment analysis could be done with greater accuracy, but they suffer from major drawback when applied to longer sentences. This paper proposes a sentiment analysis model using Long ShortTerm Memory (LSTM) based approach , which is a variant of RNNs. LSTMs are good in handling long sentence data. The model is applied to reviews collected from IMDB dataset. It is large dataset that contains 50K reviews. Out of the available reviews 50 % are used for training purpose and 50% are used for testing purpose. The model gives a training accuracy of 92% and validation accuracy of 85% which is neither an over fit nor an under fit. The overall accuracy here is 85%, which seems to be better than some of the existing techniques such as SVM with linear kernel.