Talk · International Conference on Advanced Machine Learning and Data Science · July 1, 2026
Structure-Aware Hybrid Legal RAG for Grounded Question Answering in Kazakh
Arailym Tleubayeva, Birganym Abilkhaiyr, Azamat Orazgali, Aigerim Mansurova, Aday Shomanov, Zhansaya Makhambetova
Abstract
Large language models (LLMs) show strong text generation capabilities, but their direct use in legal question answering remains limited by hallucinations, weak grounding, and unreliable citation of legal norms. These challenges are particularly relevant for Kazakh, a low-resource language with rich morphology and limited legal NLP resources. This paper presents a structure-aware Legal RAG system for grounded question answering over the legislation of the Republic of Kazakhstan. The approach combines structured legal corpus preparation, legal-aware chunking, hybrid retrieval, and citation-oriented answer generation. We compare sparse, dense, and hybrid retrieval strategies and evaluate both LLM-only and retrieval-augmented settings using Exact Match, F1, citation accuracy, support rate, hallucination rate, and legal correctness. The results show that structure-aware chunking improves retrieval quality, while retrieval augmentation improves grounding and reduces hallucinations relative to LLM-only baselines.
Related publications
- Document Segmentation in Legal Retrieval-Augmented Generation: An Empirical Study on Kazakh Legal TextsIEEE International Conference on Smart Information Systems and Technologies · 2026 · Citations: 1 · Same author: Azamat Orazgali
- Legal Eligibility Inference from Text: Constraint Extraction with Pretrained Language ModelsProcedia Computer Science · 2025 · Citations: 7 · Shared topic: Topic Modeling
- CulturalRiddles: A Multicultural Benchmark of RiddlesarXiv · 2026 · Shared topic: Topic Modeling
- A Knowledge-Augmented Multi-Agent AI System for Context-Aware Legal Reasoning and Case Outcome PredictionPreprints.org · 2026 · Shared topic: Topic Modeling