Graph-Enhanced Legal Case Retrieval for Precedent Recommendation in Turkish Jurisprudence

Companion data and code package for the article "Graph-Enhanced Legal Case Retrieval for Precedent Recommendation in Turkish Jurisprudence" (M. M. Köse, N. Kesim Cicekli, Middle East Technical University).The package contains the dataset and all experiment notebooks needed to reproduce the study, which recommends relevant precedent decisions for Turkish Court of Cassation cases using dense embeddings and a citation/similarity graph trained with GraphSAGE under a contrastive InfoNCE loss. Data: 6,011 reference (precedent) cases and 44,447 citing cases with their docket numbers and citation relationships, yielding 56,182 ground-truth citation pairs; 802,873 NER-derived case-to-law mappings; statutory article texts retrieved from the Turkish Legislation Information System; and the node and edge lists of the FAISS-built similarity network (50,458 nodes).Code: Jupyter notebooks covering the full pipeline — BERTurk NER fine-tuning, case and law embedding generation (BGE-M3, Qwen3, Jina), FAISS similarity network construction, GraphSAGE training with feature concatenation and Optuna hyperparameter optimization, and the ablation studies (GAT, HeteroGAT, HGT, graph density, shared-law edges, hub downsampling).Personal-data policy: Full decision texts are not redistributed, as they may contain personal data protected under Turkish law (KVKK, Law No. 6698). The package contains only document identifiers, public docket numbers, citation relationships, and statutory references; the released identifiers allow the full texts to be independently retrieved from the official public portal (karararama.yargitay.gov.tr) from which they originate. The NER-annotated corpus is likewise excluded, as it consists of labeled spans of raw decision text.Licensing: the datasets are released under CC BY 4.0; the notebooks under the MIT License. See README.md in this record for the file-by-file description and step-by-step reproduction instructions.