Turkish-NLI/legal_nli_TR_V1

---
license: apache-2.0
dataset_info:
features:
- name: premise
dtype: string
- name: hypothesis
dtype: string
- name: label
dtype: string
splits:
- name: train
num_bytes: 1858442640
num_examples: 474283
- name: validation
num_bytes: 18996841
num_examples: 5000
- name: test
num_bytes: 19683829
num_examples: 5000
download_size: 725637794
dataset_size: 1897123310
configs:
- config_name: default
data_files:
- split: train
path: data/train-*
- split: validation
path: data/validation-*
- split: test
path: data/test-*
task_categories:
- sentence-similarity
language:
- tr
tags:
- legal
size_categories:
- 100K<n<1M
---

# Turkish Law NLI Dataset
This dataset is derived from case files of Turkish Commercial Courts and was prepared as part of a student project to contribute to the Turkish NLP literature.

## Source Data
The dataset was created by collecting approximately 33,000 case rulings from [open sources](https://emsal.uyap.gov.tr/) using web scraping methods. The dataset includes only the "summary" sections of the case rulings, where the reason for each lawsuit is typically described.

## Data Structure and Labeling
- The dataset was adapted for sentence similarity tasks, inspired by the [SNLI dataset](https://huggingface.co/datasets/stanfordnlp/snli). The goal of this project is to develop a semantic search model for identifying relevant precedent cases in legal settings.
- This is the first version of the dataset, and future versions will incorporate additional metadata and employ more refined labeling techniques.

 
<div style="text-align: center; opacity: 0.7;">
<p style="font-style: italic;">Some sections of the Tree Structure</p>
</div>

## Labeling Methodology
To establish relationships between case files, legal articles within each case were utilized. Only commercial cases governed by the [Turkish Commercial Code (TTK)](https://www.mevzuat.gov.tr/mevzuat?MevzuatNo=6102&MevzuatTur=1&MevzuatTertip=5) are included. Articles from the TTK were aligned in a hierarchical structure, considering main and subheadings, and were transformed into a tree structure. The relationship between cases was determined by calculating distances between the articles they contain within this tree structure.

### Label Types
- **Entailment:** For each case, the 7 closest cases (with lower distances indicating closer relationships) were labeled as related.
- **Contradiction:** For each case, the 7 most distant cases were labeled as unrelated.
- **Neutral:** Each case was labeled as neutral with respect to the legal articles it contains.

## Contributors
- Mesut Demirel
- Recep Karabulut