Aspect Based Sentiment Analysis Using the Deep Learning Convolutional Neural Networks

Abstract

Sentiment Analysis is also known as Opinion Mining, is the computational study of unstructured textual information. It might be in regard to a person’s perspective, attitudes, feeling and emotions toward an event or an entity in the form of a piece of text. Sentiment analysis has become an important task for automatically classifying a piece of text as positive, negative or neutral. It helps to explore meaningful information from the data over the internet. Three important aspects are presented. The first one is creating ontologies for the extraction of semantic features. It gives effective information about domain in our case. Overall score for opinion is calculated by using these features and further we label the dataset. Second aspect involves word2vec for conversion of processed corpus. Word2vec is an unsupervised neural network, which is used to extract feature vector used by deep learning algorithm. The third and final aspect is Convolutional neural network (CNN) for training and testing. In this thesis, the proposed framework can be expressed as combination of Ontology, Word2vec and CNN. Ontology is a technique for knowledge representation and association between different entity and attributes of a specific domain. Experiments show that the use of CNN along with Ongology is an efficient approach for opinion mining. As a classifier, it achieves 88.5%, 94.3% and 81.8% in accuracy, precision and recall respectively making it more efficient compared to other state of art algorithms such as compare to SVM, Random Forest, Decision Tree, Maximum Entropy, Generalized Linear Model, Stabilized Discriminant Analysis.