VLUS06/VietOnlineNews

VietOnlineNews: Vietnamese Online News Topic Classification Dataset

Dataset Description

VietOnlineNews is a Vietnamese online news dataset constructed for the task of single-label multi-class topic classification. Each sample corresponds to one news article and is assigned exactly one main topic label through the category field.

The dataset was collected from multiple Vietnamese online news sources and processed through a data cleaning pipeline to remove incomplete records, label conflicts, duplicated content, extremely short articles, and common noise introduced during web crawling.

The main objective of this dataset is to support research and experimentation in Vietnamese Natural Language Processing (NLP), especially topic classification of Vietnamese news articles.

Dataset Summary

Data Sources

The dataset was collected from three Vietnamese online news sources:

• —VietnamNet
• —Thanh Niên
• —Tuổi Trẻ

The distribution of articles across sources follows the natural distribution of the collected data and was not artificially balanced.

Dataset Structure

The core dataset contains four main columns:

The category column is the target label used for topic classification.

Labels

The dataset contains 13 topic categories:

• —Bất động sản
• —Đời sống
• —Du lịch
• —Giáo dục
• —Giải trí
• —Khoa học công nghệ
• —Kinh doanh
• —Pháp luật
• —Sức khỏe
• —Thế giới
• —Thể thao
• —Thời sự
• —Xe

Each article belongs to exactly one category.

Dataset Splits

The cleaned dataset is divided into train, dev, and test sets using stratified splitting based on the category label.

Stratified splitting was used to preserve the label distribution across all splits.

Data Cleaning Pipeline

The dataset was processed using a multi-step cleaning pipeline.

1. Technical Normalization

Initial normalization was applied to make the raw data more consistent:

• —Unicode normalization
• —Whitespace normalization
• —Tab and newline normalization
• —Empty string detection
• —Category normalization
• —Date parsing for public_date

2. Missing Value Handling

Records were removed if they were missing at least one of the three core text fields:

• —title
• —description
• —content

No artificial text imputation was applied, since filling missing article text may distort the semantic meaning of the sample.

3. Label Conflict Removal

Label conflict occurs when the same text input is assigned to different category labels.

Records with the same combination of:

• —title
• —description
• —content

but different category values were removed entirely to preserve label consistency.

4. Duplicate Content Removal

Duplicate records were removed based on the content field.

The pipeline did not remove records solely because of duplicated title or description, since such duplication can naturally occur in news articles, especially in recurring news formats.

5. Short Article Filtering

A full_text field was temporarily created by combining:

title + description + content