Dataset Card for vietnamese-news-copus-segmented

Dataset Summary

This dataset is a refined collection of Vietnamese news articles, originally sourced from ademax/binhvq-news-corpus. It has been processed through a specialized pipeline for cleaning, normalization, and word segmentation. It is ideal for training Vietnamese Language Models (LLMs), word embeddings, or text classification tasks.

• Original Source: ademax/binhvq-news-corpus
• Language: Vietnamese (vi)
• Format: Word-segmented text (using VnCoreNLP)

Data Processing Pipeline

The dataset was constructed using the following automated pipeline:

1. HTML Cleaning: Stripped all HTML tags and boilerplate using BeautifulSoup.
2. Text Normalization:
   • Applied Unicode NFC normalization.
   • Standardized punctuation (quotes, dashes, ellipses).
   • Removed control characters and normalized whitespaces.
3. Boilerplate & Signature Removal:
   • Used Regex to remove common "See more", "Source", and "Photo by" patterns.
   • Applied heuristic rules to detect and strip journalist names, contributor info (CTV), and editorial signatures at the end of articles.
4. Length Filtering: Retained only high-quality articles with character lengths between 500 and 20,000.
5. Deduplication: Performed exact deduplication using MD5 hashing to ensure data uniqueness.
6. Word Segmentation: Segmented compound words using VnCoreNLP (e.g., trí tuệ nhân tạo becomes trí_tuệ nhân_tạo).

Dataset Structure

The dataset contains a single column:

• text (string): The final cleaned and word-segmented Vietnamese text.

Usage

You can load this dataset using the Hugging Face datasets library:

from datasets import load_dataset

dataset = load_dataset("trungbb8/vietnamese-news-copus-segmented")
print(dataset['train'][0]['text'])

Technical Specifications

• Segmentation Tool: VnCoreNLP (wseg annotator).

• Storage: Sharded at max 500MB per file for efficient streaming and downloading.

Limitations

• Heuristics: While the cleaning process is robust, some unconventional author signatures might remain or very short concluding sentences might be accidentally removed.

• Segmentation: If your model requires syllable-level input, simply replace underscores (_) with spaces.

License

This dataset is derived from the binhvq-news-corpus. Users should refer to the original source's licensing terms.

Note: This dataset was uploaded using an automated pipeline. For any inquiries, please contact the repository owner.