A Hybrid Transformer-Based Framework for Multi-Document Summarization of Turkish Legal Documents

The rapid growth in the volume and complexity of legal documents has created an urgent need for automated tools to help legal professionals efficiently extract and summarize critical information. While numerous studies have addressed this issue across various languages, Turkish legal document summarization presents unique challenges due to the language’s agglutinative structure and domain-specific terminology. This gap is particularly significant, given Turkey’s extensive legal system and the increasing volume of digital legal documents. This study represents the first comprehensive effort to address this gap by introducing a hybrid approach that combines traditional extractive techniques, including TF-IDF and TextRank, with advanced transformer-based models such as LED, Long-T5, BART-large, and GPT-3.5 Turbo. A new dataset of 2,000 Turkish civil cases, carefully curated and validated from publicly accessible legal platforms, was developed to support this research. Extractive summarization methods were evaluated based on cosine similarity, content precision, recall, and F1-score, with TF-IDF achieving a content F1-score of 0.47 using keyword-based summaries. Transformer-based models demonstrated significant improvements, with GPT-3.5 Turbo achieving the highest ROUGE scores (ROUGE-1: 55%, ROUGE-2: 35%, ROUGE-L: 42%, ROUGE-Sum: 44%), showcasing its superior ability to generate accurate and coherent summaries. By tackling the unique linguistic challenges of Turkish agglutinative morphology, this research provides the first foundation for scalable, automated tools to enhance legal professionals’ efficiency and decision-making within Turkey’s legal system.