A Context-Aware Hybrid Model for Sentiment Analysis

Abstract

The rapid growth of social media and online platforms has generated large volumes of opinion-rich text, increasing the need for automated sentiment analysis. Traditional machine learning methods can efficiently classify text but often have limited ability to capture semantic context, negation, sarcasm, and domain-specific expressions, whereas modern contextual models such as BERT provide richer representations at higher computational cost. This paper surveys research on contextaware sentiment analysis published from 2020 to 2026 and categorizes existing approaches according to their representation and classification techniques. Based on the identified trends and limitations, a hybrid opinion-mining framework is proposed that combines text preprocessing, TF-IDF statistical features, TextBlob lexical polarity, BERT contextual representations, and CNN-BiLSTM with attention mechanisms for sentiment classification and emotion recognition. The surveyed studies, datasets, reported findings, advantages, and limitations are comparatively discussed to identify research gaps. The analysis highlights the potential of combining complementary statistical, lexical, contextual, and sequential representations while also emphasizing challenges related to computational complexity, sarcasm, negation, multilingualism, and domain adaptation. The proposed framework provides a structured basis for future empirical evaluation across multiple domains.