Aspect Based Sentiment Classification with Various Feature Extraction and Selection Techniques using Machine Learning

Abstract

Aspect-based sentiment classification has various applications, such as analysing product reviews, social media sentiment analysis, or customer feedback analysis. It provides a more fine-grained understanding of sentiment by capturing the sentiment expressed towards specific aspects, enabling businesses to gain deeper insights into customer opinions and preferences. After detection of aspects, it works on Sentiment classification (SA) which focuses on determining the sentiment polarity associated with each aspect. It involves assigning sentiment labels to each aspect. This step often utilizes machine learning algorithms. A properly selected feature set is key for aspect prediction. This thesis primarily offers three systems. In the first system, a novel methodology called "two phase weighted correlation feature selection" is proposed for identifying the important elements of the items under consideration. Linguistically linked features indicate the aspects as well as the sentiments towards it. On addition, the suggested system looked at the impact of linguistic rule-based features in the aspect prediction job. Several previous techniques treated ABSA as a binary class or multiclass problem, with each review phrase predicting only one class. Since a review may discuss various characteristics of an entity, the second method suggested in this thesis uses multilabel classifiers to handle ABSA. In the third system, sentiments for extracted aspects are determined using a RNN-based approach for an end-to-end ABSA. The different machine learning algorithm provides 90% accuracy while the deep learning provides 96% average accuracy with various cross validation. The thesis presents novel approaches for each subtask that outperform existing methods. Evaluation measures such as F1-score, accuracy per label, and hamming loss are used to assess the performance of these subtasks.