Evolving optimal text clusters: A novel GA-driven framework for dynamic ensemble fusion of multi-model contextual embeddings

by Ali Sabah, Zaid Alaa

Text clustering is an essential activity in unsupervised natural language processing (NLP), and allows the automatic structure of large-scale textual collections (in natural language processing) in news classification, routing of technical questions, and text summarisation. With the emergence of contextual embedding models, such as SBERT, RoBERTa, and DistilBERT, there has been a significant improvement in the quality of clustering, with these models producing dynamic and context-sensitive representations that are more likely to reflect domain-specific semantics than the traditional word embeddings of Word2Vec and GloVe. Although the models of contextual embedding have their own advantages, they show varying performance in different domains and the current ensemble techniques propose fixed, flat fusion weights which do not utilize their complementary abilities. There is no current solution to dynamic, label-free optimisation of weight to multi-model contextual embedding fusion in unsupervised clustering which would bridge a major gap between fixed integrative approaches and adaptive, domain-accommodative model combination. This paper presents a new Genetic Algorithm (GA)-based ensemble model that dynamically optimizes the best fusion weights of SBERT, RoBERTa, and DistilBERT embeddings without ground-truth labels. The hypothesis is that evolutionary optimisation can discover domain-adaptive weight sets that outperform standalone models as well as fixed ensemble baselines on linguistically heterogeneous datasets. The framework uses L2-normalisation and topology in which the topology is a sum-to-one constraint so that the models can be fairly integrated, and a composite fitness measure based on Silhouette Score, Adjusted Rand Index, and Topic Coherence to direct the weight evolution through tournament selection, uniform crossover, and Gaussian mutation. The proposed framework on three heterogeneous benchmarks, AG News, 20 Newsgroups, and Stack Overflow, has Silhouette Score improvements of +14, 16 percent, and Topic Coherence gains of +17 percent, all statistically significant at p < 0.05. Evolved weights can be completely understood, depicting domain-specific model roles, which guide future architecture choice. The framework is suggested as a scalable, modular system to practical unsupervised NLP problems, and its architecture can be easily extended to new transformer systems and semi-supervised systems.