Context-Aware Hybrid Text Generation Integrating Bidirectional Long Short-Term Memory Sequencing with Semantic Clustering

Abstract

Natural language generation systems sometimes struggle to model long-range semantic connections and maintain contextual consistency, especially when applied to linguistically sophisticated literary corpora. Traditional recurrent neural architectures are good at modeling sequential patterns but typically fail to preserve higher-level thematic and stylistic information in text production. The current paper proposes a semantic-aware hybrid framework based on Word2Vec embedding representations, ++K-Means semantic clustering, and Bidirectional Long Short-Term Memory (Bi-LSTM) sequence learning to improve contextual coherence and next-word prediction performance. In the proposed architecture, the model learns to obtain semantic context vectors from clustered embedding spaces and to fuse them with sequential hidden representations for better language modeling. This study examines the system on three benchmark datasets from English corpora of both literary and general domains: the Nietzsche corpus, Shakespeare plays and WikiText-2. Experimental results show that the proposed semantic-aware recurrent architecture consistently outperforms the standard statistical and neural baseline models. The model achieves prediction accuracies of 67.4%, 61.3%, and 63.1% on the Nietzsche, Shakespeare, and WikiText-2 datasets, respectively, while reducing perplexity values and enhancing linguistic coherence. A more detailed analysis of the robustness test, semantic error evaluation, and ablation experiments confirm that semantic clustering effectively can improve contextual consistency, stylistic preservation, and semantic continuity. The results demonstrate that combining clustering-based semantic abstractions with recurrent sequence modeling is an effective, computationally lightweight approach to context-aware text synthesis for both literary and general-domain applications.