This paper introduces a novel sentiment analysis framework that synergistically integrates statistical word association measures with deep learning architectures, such as Long Short-Term Memory (LSTM), Bidirectional LSTM (BiLSTM), and Recurrent Neural Networks (RNN), for improving the granularity and accuracy of sentiment classification. Our innovation is twofold: (1) extract fine-grained semantic relationships between words using Pointwise Mutual Information (PMI), extending over the large limitations of the existing bag-of-words method in terms of the failure to detect context-dependent sentiment cues, and (2) an extended Semantic Orientation (SO) metric that captures a more accurate intensity measure of sentiment by considering syntactic dependencies and domain-specific polarity shifts. These improvements allow the model to be more sensitive when distinguishing between positive, negative, and neutral aspects, especially when dealing with noisier domains such as microblogging systems. By coupling these statistical augmentations with deep learning models, our framework delivers superior contextual knowledge, using not only long-range dependencies but also localized semantic structures. Experimental results on large-scale social media datasets demonstrate the suitability of our approach in customer feedback analysis, real-time social media monitoring, and automated content moderation. This work provides evidence about the ability of hybrid methodologies that combine statistical linguistics and neural networks to improve the robustness of sentiment analysis.