UNRAVELING AMBIGUOUS SENTIMENT CONTEXTS: A NOVEL APPROACH USING GAUSSIAN DISTRIBUTION-BASED LONG SHORT-TERM MEMORY (GD-LSTM) FOR IMPROVED SENTIMENT ANALYSIS

Abstract

Sentiment analysis, a fundamental task in natural language processing, involves classifying text data into sentiment categories such as positive, negative, or neutral. However, conventional sentiment analysis methods encounter challenges in accurately capturing nuanced linguistic features, contextual dependencies, and ambiguous expressions, leading to suboptimal performance. To mitigate these challenges, this study introduces a novel approach utilizing Gaussian Distribution-based Long Short-Term Memory (GD-LSTM) networks. Unlike standard LSTM models, GD-LSTM incorporates uncertainty estimation through Gaussian distributions, enabling the model to represent and manage uncertainty in sentiment predictions quantitatively. By integrating uncertainty modelling, GD-LSTM enhances its capacity to capture the inherent uncertainty in sentiment analysis tasks, facilitating more robust and reliable sentiment predictions, particularly in contexts characterized by ambiguity or contextual variability. Operationally, GD-LSTM leverages its uncertainty estimates to optimize sentiment prediction by effectively balancing model confidence with uncertainty awareness. This mechanism enables GD-LSTM to make informed decisions even in challenging scenarios where sentiment polarity may be ambiguous or contextually dependent. To assess the efficacy of GD-LSTM, experiments were conducted using the widely recognized Amazon product review dataset, which presents a diverse range of sentiment analysis challenges. The experimental findings underscored GD-LSTM's superiority over traditional LSTM models, showcasing its ability to achieve state-of-the-art performance in accurately capturing ambiguous sentiment contexts and enhancing sentiment prediction accuracy across various textual data domains.