An enhanced MNB based model for explicit and hidden sentiment classification in imbalanced datasets

Abstract

Sentiment polarity classification (either explicit or hidden) is the process by which information can be extracted to be analysed as positive or negative opinion. Much work on supervised machine learning based sentiment classification has been done considering balanced datasets. However, due to the imbalanced nature of data distribution, sentiment classification becomes a complex task that requires investigating more efficient approach, especially for hidden sentiment. Multinomial Naïve Bayes (MNB) algorithm is one of the most widely used methods in this field, due to its computational efficiency and relatively good predictive performance. However, this classifier performs poorly on imbalanced datasets. In this paper, we propose transformation in MNB prior and conditional probabilities designed to handle explicit and hidden sentiment classification for imbalanced data. To support MNB, we introduce an original frequency model based on using synonym and antonym semantic relations. Our approach is empirically evaluated according to: (1) its impact on MNB training data quality for minority classes, (2) its comparison to similar works involving MNB for imbalanced datasets and (3) its comparison to commonly used classifiers for sentiment classification. The experimental results show that our model improves training data quality and therefore helps MNB boost its performance to achieve the best results for explicit and hidden imbalanced datasets.