Towards Human Thought Classification: A Sentiment Analysis Framework using Active and Deep Learning

Understanding and classifying human thoughts is essential to take sentiment analysis beyond just detecting whether something is positive or negative. Most existing methods miss the complex nature of human thinking and reduce it to basic labels like "good" or "bad." This limits their usefulness in important areas like mental health analysis, education, and behavior modeling. Unlike existing works that limit sentiment to binary or ternary classes, our work introduces a four-label thought classification schema, supported by an actively learned, real-world dataset and a performance-optimized BiLSTM pipeline. In this research, we propose a detailed system that can classify thoughts into four categories: positive, negative, necessary, and peripheral. To reflect real human thinking, we created a new manually labeled dataset using conversations from online platforms. Introducing a new system that intelligently blends active learning—allowing the model to focus on tougher cases—alongside classic machine learning strategies and current deep learning methods. This approach cuts down on manual labeling effort while still delivering strong accuracy. Our tests indicate that BiLSTM-based models, particularly enhanced with TF-IDF features and careful data preparation, tend to outperform alternative models across various performance measures. Beyond just the promising results, the system offers real-world advantages, potentially aiding in mental health monitoring. Ultimately, by prioritizing thought classification and leveraging smart sampling with deep learning, this research presents an innovative, scalable method for sentiment analysis.