| Kübra Kovayçin, Ece İrem Şişer, Derda Sina Günay, Yusuf Evren Aykaç, Refik Samet Section-Aware Retrieval for Turkish Case Law: A Case Study on Scanned Court Decisions |
|---|
| Abstract. Finding relevant precedents often needs more than topical overlap: operative facts and judicial reasoning can be decisive. In Turkey, court decisions are often long and frequently shared as scanned (imagebased) PDFs, making extraction and indexing harder. We present a section-aware retrieval case study on 1,000 Turkish court decisions about social media–related disputes. Using a vision–language model, each PDF is transcribed and organized into Facts, Reasoning, and Verdict segments; statute/article mentions are also extracted as lexical anchors. We compare a lexical baseline (BM25) with multilingual and legal biencoders adapted via parameter-efficient fine-tuning, and we test lexical–dense hybrid fusion, including a simple section-aware hybrid that scores sections separately under fixed token budgets. Because expert-labeled precedent links are not available, we use a self-alignment known-item diagnostic: an LLM generates a short, query-like summary, and the system tries to retrieve the decision it came from; full-document BM25 is included as a reference baseline. Under this proxy, BM25 over Reasoning is the strongest single-section index (R@1=0.744), and section-aware hybrid fusion with MPNet reaches R@1=0.780 and MRR@10=0.810, outperforming naive fusion while preserving section-level interpretability. This proxy setup is intended as a diagnostic rather than a full relevance benchmark; retrieval-time scoring is deterministic, using BM25 and cosine similarity over fixed embeddings. |
| Keywords: Legal information retrieval, precedent retrieval, section-aware retrieval, BM25, neural bi-encoder models, sparse–dense hybrid fusion, large language models (LLMs), Turkish case law |
 Download PDF |
| DOI: https://doi.org/10.54381/itta2026.3.03 |