Datasets:

gradientflow
/

legal-ner

language:
  - sr
license: cc-by-4.0
task_categories:
  - token-classification
tags:
  - ner
  - legal
  - conllu
dataset_info:
  features:
    - name: id
      dtype: string
    - name: doc_id
      dtype: string
    - name: tokens
      sequence: string
    - name: ner_tags
      sequence:
        class_label:
          names:
            '0': B-ADR
            '1': B-COM
            '2': B-CONTACT
            '3': B-COURT
            '4': B-DATE
            '5': B-IDCOM
            '6': B-IDOTH
            '7': B-IDPER
            '8': B-IDTAX
            '9': B-INST
            '10': B-LAW
            '11': B-MISC
            '12': B-MONEY
            '13': B-NUMACC
            '14': B-NUMCAR
            '15': B-NUMDOC
            '16': B-NUMPLOT
            '17': B-ORGOTH
            '18': B-PER
            '19': B-REF
            '20': B-TOP
            '21': I-ADR
            '22': I-COM
            '23': I-COURT
            '24': I-DATE
            '25': I-INST
            '26': I-LAW
            '27': I-MISC
            '28': I-MONEY
            '29': I-NUMCAR
            '30': I-ORGOTH
            '31': I-PER
            '32': I-REF
            '33': I-TOP
            '34': O

Dataset Card for Legal NER dataset

Dataset Description

This dataset is based on COMtext.SR.legal, the first corpus of legal-administrative texts in Serbian that has been manually annotated for Named Entities (NER) according to the IOB2 standard.

The corpus was created during 2023. The selection of representative legal texts of various types (contracts, judgments, conclusions, decisions, requests, appeals, rulebooks, laws, decrees, statutes, etc.) was conducted with the assistance of the lawyers.

The corpus includes 79 documents, comprising 4,762 sentences or 105,470 tokens.

NER Statistics

Out of a total of 105,470 tokens, 14,113 (13.4%) belong to named entities.

Systematization and frequency of entity types:

Category	Subcategory	Tag	Count / Percentage	Avg. Length in Tokens
Persons		PER	694 (19.2%)	1.89
Locations	Toponyms	TOP	294 (8.1%)	1.24
	Addresses	ADR	203 (5.6%)	6.96
Organizations	Courts	COURT	148 (4.1%)	3.59
	Institutions	INST	395 (10.9%)	2.75
	Companies	COM	337 (9.3%)	2.28
	Other Organizations	ORGOTH	97 (2.7%)	2.27
Legal Docs	General Legal Acts	LAW	395 (10.9%)	9.17
	Individual Legal Acts	REF	227 (6.3%)	10.52
Sensitive Data	Personal ID (JMBG)	IDPER	21 (0.6%)	1.0
	Company Reg. No.	IDCOM	33 (0.9%)	1.0
	Tax ID (PIB)	IDTAX	16 (0.4%)	1.0
	Bank Account No.	NUMACC	6 (0.2%)	1.0
	ID Card/Passport No.	NUMDOC	9 (0.2%)	1.0
	Car Plate/Chassis	NUMCAR	6 (0.2%)	2.67
	Cadastral Plot No.	NUMPLOT	67 (1.9%)	1.0
	Other IDs	IDOTH	18 (0.5%)	1.0
	Contact (Email, Phone)	CONTACT	8 (0.2%)	1.0
Time/Money	Dates	DATE	352 (9.7%)	4.1
	Monetary Amounts	MONEY	246 (6.8%)	2.32
Other		MISC	39 (1.1%)	5.0

Structure & Usage

The data has been converted from CoNLL-U format to Hugging Face columnar format.

Train/Test Split

The split into training and testing sets was performed using the GroupShuffleSplit strategy on the doc_id attribute. This guarantees that sentences from the same document do not appear simultaneously in both sets, preventing data leakage and ensuring realistic model evaluation.

Usage Example

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("gradientflow/legal-ner")

# Display the first example from the train set
print(dataset['train'][0])

# Display all available tags
print(dataset['train'].features['ner_tags'].feature.names)

Source Data

The original annotated corpus is available on GitHub:

COMtext.SR