nezahatkorkmaz/turkish-mevzuat-bert

SentenceTransformer for Turkish Legal Texts

This is a sentence-transformers model fine-tuned from atasoglu/turkish-base-bert-uncased-mean-nli-stsb-tr. It is specifically optimized for semantic search and sentence similarity in formal Turkish legal documents such as regulations, directives, and academic institution bylaws.

Model Overview

• Model Type: Sentence Transformer
• Base Model: atasoglu/turkish-base-bert-uncased-mean-nli-stsb-tr
• Max Sequence Length: 256 tokens
• Output Dimension: 768-d vectors
• Loss Function: CosineSimilarityLoss

Use Case: Turkish Legal Semantic Search

The model was trained on yusufbaykaloglu/turkish-university-mevzuat, a curated dataset of official Turkish university regulations. Sentence pairs were labeled with semantic similarity scores and sampled from structured legal sections.

It was specifically developed to:

• Improve semantic search across long and repetitive legal texts
• Enhance sentence-level retrieval in regulation and policy documents
• Offer a meaningful alternative to keyword-based Elasticsearch queries

Training Details

• Training Samples: 9,000 sentence pairs
• Tokenizer Max Length: 256 tokens
• Epochs: 4
• Batch Size: 16
• Optimizer: AdamW
• Evaluation Strategy: Steps (every 1000 steps)
• Final Spearman Cosine: 0.6625

Sample Data Format

sentence_0	sentence_1	label
"Yönetim Kurulu görevleri..."	"Müdürün önerisi ve Rektörün onayı..."	0.8
"Merkez faaliyetleri..."	"Danışma Kurulu görevleri..."	0.1

Example Usage

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("nezahatkorkmaz/mevzuat-bert-turkish")

sentences = [
    "(1) İhalesi yapılacak her iş için bir onay belgesi hazırlanır...",
    "Yönetim Kurulunun görevleri"
]
embeddings = model.encode(sentences)

Citation

@inproceedings{reimers-2019-sentence-bert,
  title={Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks},
  author={Reimers, Nils and Gurevych, Iryna},
  booktitle={EMNLP},
  year={2019},
  url={https://arxiv.org/abs/1908.10084}
}

License

This model inherits the license of the base model: CC BY-SA 4.0.

---

Developed by Nezahat Korkmaz.