O-Rag Ontology-Enhanced Retrieval-Augmented Generation For The Uzbek Legal Domain

NLP tasks that depend on extra knowledge retrieval-based generation (RA) architectures have improved the performance of knowledge-intensive models on knowledge-based NLP tasks, but current models tend to be unable to cope with the semantic challenges of specialized tasks and morphologically rich low-resource languages such as Uzbek. This is especially a major limitation within the field of law where accuracy and contextual comprehension is of utmost importance. The paper presents a new framework O-RAG, which is specifically developed on the basis of the Uzbek legal sphere. O-RAG combines an additional custom legal ontology to drive a hybrid retrieval process, where a new ontological re-ranking algorithm gives conceptually relevant legal articles a higher rank than articles that are simply lexically similar. Our model was tested on a new question-answering dataset that we developed based on the official legislative database in Uzbekistan, lex.uz. As shown by the experimental findings, the O-RAG is by far superior to the standard RAG baselines in all measures, making it higher in Retrieval Precision (Hit Rate) and higher in Citation Accuracy-a key indicator to use in a legal setting. Our paper confirms the usefulness of integrating domain knowledge such as structured knowledge into neural retrieval models in constructing more reliable and accurate AI systems in high stakes settings.

Перевод пока недоступен