Document Type : Original/Review Paper

Authors

• Mohammad Rahmanimanesh ¹
• mina karimian ²
• Farzin Yaghmaee ²

¹ Faculty of Electrical and Computer Engineering, Semnan University, Semnan, Iran

² Faculty of Electrical and Computer Engineering, Semnan University, Semnan, Iran.

Abstract

Extracting meaningful insights from unstructured text data remains a significant challenge, particularly for Persian social media comments. This paper proposes a novel, multi-stage unsupervised framework for extracting hidden thematic patterns without requiring labeled data for training. The core innovation lies in a knowledge-based representation that enriches the TF-IDF matrix with the ConceptNet knowledge graph. This approach enables concept-based analysis by identifying synonymous and semantically related terms, transcending surface-level lexical matching. To mitigate noise and the "curse of dimensionality," Latent Semantic Analysis (LSA) is applied, followed by a Spherical k-means clustering algorithm for optimal semantic alignment. Finally, cluster topics are automatically extracted and interpreted through a Conceptual Analysis (CA) layer powered by class-based TF-IDF (c-TF-IDF), ensuring high interpretability without requiring manual topic annotation. The framework was validated on a domain-specific corpus of 512 unique Persian user comments curated from Instagram and specialized car websites. Quantitative evaluation shows that the knowledge-enriched model achieves a Silhouette Score of 0.7359. To provide extrinsic validation, the discovered clusters were mapped to human-annotated labels, yielding a classification accuracy of 72.5%. The findings successfully identified and quantified four core consumer concerns: Engine (63.1%), Price (14.8%), Aesthetics (7.4%), and Performance (14.7%). These results offer strategic insights for manufacturers and, given the computationally lightweight and knowledge-driven design of the proposed framework, suggest potential for extension to larger corpora and other low-resource language domains.

Keywords

• Persian Text Processing
• User Preferences Extraction
• Knowledge Graph-Based Semantic Model
• K-means Clustering

Main Subjects

• Document and Text Processing