Ethio-PolicyNet: A Hybrid Neuro-Symbolic Framework for Explainable Policy Document Classification in Low- Resource Settings

Abstract Background: The digital transformation of governance in Ethiopia is hindered by a massive volume of unstructured, multilingual policy documents. Traditional Natural Language Processing (NLP) approaches often struggle with the morphological complexity of Amharic and the nuanced, domain-specific logic required for accurate policy analysis. Objective: This paper proposes Ethio-PolicyNet, a hybrid neuro-symbolic framework designed to automate the classification and semantic enrichment of Ethiopian policy documents. The goal is to improve classification accuracy in low-resource settings while providing humaninterpretable explanations for model decisions. Methodology: We introduce a dual-stream architecture that integrates sub-symbolic neural patterns with formal symbolic knowledge. The system utilizes a multilingual transformer (mBERT/XLM-R) as a neural encoder and the Ethiopian Policy Ontology (EPO)a domainspecific knowledge base containing 1,240 axiomsas a symbolic scaffold. A novel Gated Fusion Mechanism is employed to dynamically weight these streams, allowing the model to rely on symbolic logic when linguistic data is sparse or ambiguous. Results: Experimental results on a curated corpus of 2,400 policy segments demonstrate that Ethio-PolicyNet achieves an overall F1-score of 92.4%, significantly outperforming baseline transformer models (87.5%) and classical machine learning approaches (71.2%). Most notably, the model stabilizes performance for Amharic text, achieving a 91.8% F1-score, effectively “closing the gap” between high-resource and low-resource language performance. Conclusion: By bridging the gap between deep learning and formal knowledge representation, Ethio-PolicyNet provides a robust, transparent tool for public-sector decision-makers. The framework not only automates administrative workflows but also enhances the accountability of AI-driven policy informatics through ontology-grounded explainability.