A Multi-Model Framework for Autonomous Schema Discovery and Hybrid Natural Language Generation-Driven Augmented Business Intelligence

Randy Permana (Universitas Putra Indonesia YPTK Padang)
Sarjon Defit (Universitas Putra Indonesia YPTK Padang)
Gunadi Widi Nurcahyo (Universitas Putra Indonesia YPTK Padang)

Journal of Applied Data Sciences
Vol 7, No 3: September 2026

Traditional Business Intelligence (BI) frameworks rely heavily on data professionals to manually map relational metadata into structured star schemas during the ETL process, creating a significant operational bottleneck in data preparation and downstream interpretation of insights. To address these limitations, this study introduces an augmented analytics framework for autonomous schema discovery and hybrid Natural Language Generation (NLG) driven Augmented BI. In the data representation layer, a weighted hybrid feature fusion mechanism combines structural database metadata with contextual text embeddings produced by a pre-trained Sentence Transformer (all-MiniLM-L6-v2). In the multi-model machine learning layer, a multi-paradigm execution engine combines unsupervised geometric clustering models (K-Means, K-Medoids, DBSCAN) and supervised classifiers (SVM, Random Forest), and performance is evaluated using Leave-One-Out Cross-Validation (LOOCV). The resulting schemas are then dynamically projected into an in-memory OLAP cube. At the downstream insight interpretation layer, a Hybrid NLG engine combines a context-aware, rule-based router with an autoregressive generative decoding mechanism to autonomously produce adaptive, actionable business commentaries triggered by user-driven OLAP exploration states. Experimental results demonstrate that applying linear semantic scaling optimization (α) substantially mitigates statistical semantic blindness and protects the framework from structural schema misclassification. On the E-Commerce dataset, the proposed K-Medoids+Semantic configuration demonstrated topological superiority, achieving a peak Silhouette Score of 0.611 and a compressed Davies-Bouldin Index of 0.514. Meanwhile, on the high-dimensional Superstore dataset, the pipeline maintained high functional flexibility, stabilizing overall classification accuracy up to 94.74%. Furthermore, the downstream Hybrid NLG engine (Template+Generative) demonstrated high factual integrity and linguistic flexibility, achieving a ROUGE-1 score of 0.85, a ROUGE-2 score of 0.82, and a BLEU score of 0.15. This research provides ABI frameworks that enable accelerated executive decision-making through seamless, data-to-insight automation.