alakxender/dhivehi-summaries

Dhivehi Text Summarization Dataset

Dataset Description

This dataset contains Dhivehi (Maldivian) text summarization pairs, consisting of original news articles and their corresponding summaries. The dataset is designed to support the development of Dhivehi text summarization models and advance NLP research for low-resource languages.

Dataset Summary

• Language: Dhivehi (dv)
• Task: Text Summarization
• Domain: News Articles
• Format: Original content paired with summaries
• Split: 80% train, 20% test

Languages

The dataset contains text in Dhivehi (Divehi), the official language of the Maldives, written in the Thaana script.

Dataset Structure

Data Instances

Each instance contains:

• text: The full original text/article in Dhivehi
• summary: A concise summary of the original content in Dhivehi

Example:

{
    "text": "ކޮންމެ މީހަކަށްވެސް ނުރަނގަޅު ކަމެއް ކުރެވިދާނެއެވެ. އެކުގައި އުޅެން އެންމެ ނުރައްކައުތެރި މީހުންނަކީ ކުރެވޭ ނުރަނގަޅު ކަންކަން ވިސްނައިދިނުމުން އިޞްލާޙްނުކުރާ މީހުންނެވެ. އެމީހުން ކުރާ ނުބައި ޢަމަލުކުރުންމަތީގައި ދެމިތިބޭ މީހުންނެވެ. ބައެއް މީހުންނަށް ވިސްނެނީ އެފަދަ މީހުންނާއިއެކު އުޅުމަކީ އެއްވެސް މައްސަލައެއް ނޫންކަމަށެވެ...",
    "summary": "އެއްވެސް މީހަކީ ފުރިހަމަ މީހެއް ނޫނެވެ. ނަމަވެސް، ކުށްތަކަށް އިޞްލާޙް ނުވެ ދެމިތިބޭ މީހުންނާއެކު.."
}

Data Fields

• text (string): The source text that needs to be summarized
• summary (string): Human-written summary of the original content

Data Splits

Split	Number of Examples
train	~80% of total
test	~20% of total

Length Analysis

The dataset contains summaries of varying lengths. The length of the summaries is as follows:

Total Entries : 56,133

Length Statistics

• Mean Length: 525.9 characters
• Median Length: 498.0 characters
• Min Length: 3 characters
• Max Length: 7,702 characters
• Standard Deviation: 198.5

Category Distribution

Category	Range (chars)	Count	Percentage	Description
Very Short	0-40	59	0.1%	keyword/headline
Short	41-120	54	0.1%	short highlight
Medium	121-300	2,475	4.4%	short paragraph
Long	301-600	39,516	70.4%	detailed paragraph
Extended	601+	14,029	25.0%	extended summary

Dataset Creation

Source Data

The dataset consists of Dhivehi news articles and their corresponding summaries, collected and processed to ensure quality and consistency.

Data Processing

The data underwent several preprocessing steps:

1. Validation: Ensured both original content and summaries are non-empty
2. Character Cleaning: Removed unwanted Unicode characters (RTL marks, zero-width characters)
3. Content Filtering: Removed advertisement text and promotional content
4. Whitespace Normalization: Cleaned excessive whitespace and formatting issues

Usage

Loading the Dataset

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("alakxender/dhivehi-summaries")

# Access train and test splits
train_data = dataset['train']
test_data = dataset['test']

# Example usage
print(f"Training examples: {len(train_data)}")
print(f"Test examples: {len(test_data)}")

Use Cases

This dataset can be used for:

• Text Summarization: Training and evaluating Dhivehi summarization models
• Language Model Fine-tuning: Adapting existing models for Dhivehi text generation
• Cross-lingual Research: Studying summarization across different writing systems
• Low-resource NLP: Advancing techniques for under-resourced languages

Limitations

• Domain Specificity: Primarily contains news content, may not generalize to other domains
• Size: As a low-resource language dataset, the size may be limited compared to major languages
• Temporal Coverage: Content may be concentrated in specific time periods
• Length of summaries: Some summaries may be longer depending on the use case. Additional processing may be required to standardize or shorten them.