📝 Abstract:
Legal professionals in Turkey face a growing challenge in efficiently locating rel- evant precedent cases from an ever-expanding corpus of judicial decisions. While graph-enhanced retrieval systems have demonstrated strong results in Common Law jurisdictions, no such system exists for Turkish Civil Law, where judicial reason- ing is fundamentally anchored in codified statutes rather than accumulated case-law precedents. This thesis proposes the first end-to-end, graph-enhanced precedent case recommendation system designed explicitly for the Turkish legal framework. The proposed methodology operates as a three-phase pipeline. First, a novel Turk- ish legal corpus comprising 50,458 judicial decisions from the Court of Cassation is constructed from scratch, and a fine-tuned BERTurk model is deployed for Named Entity Recognition to automatically extract cited legislation and precedent references, achieving a weighted F1-Score of 0.95 under relaxed matching. Second, dense em- beddings are generated using multilingual models, and a feature concatenation strat- egy enriches each case representation with mean-pooled statutory law embeddings to inject statutory context into the node features. Third, a GraphSAGE model is trained on the constructed similarity network using a contrastive InfoNCE loss function. vExperimental results indicate that integrating GraphSAGE with BGE-M3 embeddings and feature concatenation achieves a Hit@1 of 65.90%, a Hit@10 of 94.20%, and an MRR of 0.7733, representing improvements exceeding 50 percentage points over standalone embedding baselines. Comprehensive ablation studies further validate the architectural choices, suggesting that GraphSAGE consistently outperforms alterna- tive architectures including GAT, HeteroGAT, and HGT.