A Multi-Feature Fusion Framework for Sentiment Analysis Based on Textual and Affective Signals

Abstract

Sentiment analysis of social media content, particularly on platforms like Twitter, presents significant challenges due to the informal, brief, and context-dependent nature of user-generated text. Traditional lexicon-based and shallow machine learning approaches often fail to capture nuanced sentiment expressions, especially in the presence of slang, abbreviations, sarcasm, and emotionally charged language. To address these limitations, this paper proposes a novel tri-stream feature fusion framework that integrates contextual semantics, sequential dependencies, and affective signals for robust sentiment classification. The framework employs RoBERTa to extract rich contextual embeddings, Bidirectional Long Short-Term Memory (BiLSTM) networks to capture word-order and temporal patterns, and lexicon-based emotion vectors to enhance emotional cue detection. These heterogeneous features are concatenated at the representation level to form a comprehensive feature space, which is subsequently used to predict sentiment polarity via a fully connected neural network classifier. Extensive experiments conducted on the Sentiment140 dataset, comprising 1.6 million labeled tweets, demonstrate that the proposed approach significantly outperforms conventional baselines and recent hybrid models, achieving an accuracy of 92.1%. Additionally, ablation studies and misclassification analyses reveal each feature stream’s complementary contributions and highlight challenges in detecting sarcasm and implicit sentiment. Future work will integrate sarcasm-aware components and external knowledge sources to further enhance model interpretability and robustness.