Abstract
The major challenge manifesting in resource-efficient named entity recognition for Classical Arabic may be attributed to the language’s rich morphology, orthographic variation, and the limitation in computing budgets. Thus, this study develops and proposes a hybrid approach that is compact, integrating linguistically informed rules, genetic-algorithm (GA) feature selection, and a multinomial Naive Bayes tagger. The system is exposed to CANER with leakage-controlled splits and documented statistical procedures thereafter achieving micro-F1 93.0% (with a precision of 93.3% and recall of 92.7%), with macro-F1 88.7%. Based on per-class analysis, boundary-aware rules mainly increased precision for frequent entities, namely PER and LOC. On the other hand, GA-controlled sparsity determined recall in minority categories, as in SECT and DAY. The combined pipeline was confirmed to produce a balanced precision-recall profile compared to individual components, along with significant minimized feature dimensionality and training time. Notably, new transformer baselines are untrained owing to limitations in computations and data, but rather the contribution is generatable and interpretable reference baseline for Classical Arabic. This has transparent implementation details and evaluation safeguards for bringing about replication and accurate comparison. Even though higher macro-F1 is achieved by transformer systems, based on the findings, a meticulously engineered hybrid can produce competitive effectiveness in the face of limited settings and valuable diagnostic views through interpretative error patterns. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1038/s41598-025-30171-6.