Abstract
Aim: The proposed Filter Based Sentiment Feature Selection (FBSFS) model focuses on to improve the performance of Sentiment Learning (SL) by selecting the most relevant sentiment features from text reviews using feature selection methods at document level.
Method: Sentiment Learning is applied at the document level for classifying text reviews into two categories either positive or negative. The key sentiment features adjectives (ADJ), adverbs (ADV), and verbs (VRB) which are essential for sentiment analysis, are extracted from text document using the WordNet dictionary. Feature selection is performed by applying four different algorithms: Information Gain, Correlation, Gini Index, and Chi-Square. These algorithms help identify the most significant features that contribute to sentiment classification. The selected features are then fed into a Back Propagation Deep Learning (BPDL) classification model for sentiment analysis.
Result: The experimental findings show that the proposed model achieved higher accuracy of 91.15% using Correlation feature selection. This accuracy signifies the effectiveness of the proposed model in classifying text reviews, outperforming other methods in terms of sentiment feature selection and classification.
Conclusion: The proposed model enhances the performance of sentiment learning by selecting the most relevant sentiment features, particularly those extracted from adjectives, adverbs, and verbs, and combining them with BPDL. The FBSFS model as a robust tool for sentiment classification.

Key words: Back Propagation, Chi Square, Correlation, Deep Learning, Feature Selection