Intelligent Document Clustering for Sinhala Text: Combining LLM-based NER with Fuzzy Clustering Techniques

The landscape of Sinhala language processing presents significant challenges in natural language understanding, particularly in named entity recognition and document clustering. Addressing these challenges, a novel hybrid approach emerges, combining a custom-trained Spacy Convolutional Neural Network with the LLaMA-3.1-Sonar large language model for comprehensive named entity extraction. The methodology integrates advanced entity recognition techniques, identifying complex entity classes including Person, Location, Organization, Date, Time, and a hierarchical event classification with specialized subcategories for Political, Sports, Economic, and Cultural events. Leveraging Fuzzy C-Means clustering, the framework dynamically determines optimal cluster configurations through sophisticated elbow method analysis, evaluating multiple performance metrics simultaneously. A semantic cluster naming technique employs the LLaMA model to generate meaningful, bilingual cluster labels, transforming traditional clustering by providing interpretable context-sensitive identifiers in both Sinhala and English. The approach addresses the unique linguistic nuances of the Sinhala language, offering a robust framework for advanced text analysis and bridging existing technological gaps in low-resource language processing.