Dynamic weight adaptation in soft voting for emotion detection using neural networks

Abstract

Confirming elevated accuracy and speed in multi-label automatic emotion classification endures to pose extensive challenges. Old-style machine learning (ML) models are broadly used for this. However, large-scale fast embryonic textual information often obstructs their performance. Deep learning (DL) models resolve the former problem efficiently, but fine-tuning the hyperparameter entails a lot of work and experience. Ensemble learning practices offer enhanced accuracy, but classical soft voting classifiers with static weights fall short to adapt effectively to diverse data traits. To tackle this limitation, this study proposes a novel ensemble framework that employs a neural network (NN) based dynamic weight adaptation within a soft voting classifier. The model dynamically adjusts the weights of core ML classifiers based on their real-time predictive likelihood and performance statistics. This adaptive weighting suggestively enhances the model’s ability in detecting nuanced emotional expressions in text, improving responsiveness and generalization. Comprehensive experiments conducted on yardstick emotion dataset demonstrate that proposed integration of NN driven adaptive weighting within an ensemble framework outpaces traditional approaches, capturing an overall classification accuracy of approximately 98% thus offering a scalable and robust solution for real-world sentiment analysis applications.