szzs1693/nlp-thaipbs-news

Thai PBS News Articles

A dataset of Thai-language news articles scraped from Thai PBS (Thailand's public broadcasting service), covering three topic categories: politics, economy, and social affairs.

Dataset Overview

• Source: Thai PBS (thaipbs.or.th), collected from the public news website (https://www.thaipbs.or.th/news)
• Total records: 1,200 articles
• Categories: 3 (politics, economy, social)
• Date coverage: 31 Aug 2025 – 2 Mar 2026
• Language: Thai (th)
• Split: Single file — no official train/validation/test split

Dataset Structure

Categories

Thai name	English	Articles	Earliest article	Latest article
การเมือง	Politics	400	2025-10-28	2026-03-02
เศรษฐกิจ	Economy	400	2025-09-01	2026-02-27
สังคม	Social	400	2025-08-31	2026-02-27

Field Schema

Each record in dataset.jsonl is a JSON object with the following fields:

Field	Type	Description
url	string	Canonical article URL (e.g. https://www.thaipbs.or.th/news/content/356852)
title	string	Article headline in Thai
categoryName	string	Category label in Thai: การเมือง, เศรษฐกิจ, or สังคม
abstract	string	Lead summary paragraph in Thai; empty string for ~10 % of articles (115/1,200)
publishedTime	string	ISO 8601 datetime with +07:00 (ICT) offset, e.g. 2025-09-23T17:01:11+07:00
tags	array of strings	Editorial keyword tags in Thai
content_html	string	Article body as cleaned HTML (see Processing Notes)
content_text	string	Plain text extracted from content_html

Example Record (abbreviated)

{
  "url": "https://www.thaipbs.or.th/news/content/356852",
  "title": "สรุปราคาทองคำ 23 ก.ย.2568 ปิดตลาดวันนี้ บวก 1,000 บาท",
  "categoryName": "เศรษฐกิจ",
  "abstract": "ราคา ทองคำ ปิดตลาดผันผวน 14 ครั้ง บวกแรง 1,000 บาท...",
  "publishedTime": "2025-09-23T17:01:11+07:00",
  "tags": ["ราคาทองวันนี้", "ราคาทองวันนี้ 1 สลึง ขาย", "ราคาทองวันนี้ เยาวราช"],
  "content_html": "<p>วันนี้ ( 23 ก.ย.2568) ราคาทองโลก...</p>",
  "content_text": "วันนี้ ( 23 ก.ย.2568) ราคาทองโลก..."
}

Content Statistics

Field	Mean	Median	Min	Max
content_text chars	2,801	2,343	315	18,077
abstract chars	162	174	0	255
tags count	10	10	0	24

No records have null values; abstract is an empty string for 115 articles (~10%).

Data Processing Notes

The following cleaning steps were applied to the raw HTML before saving:

• Tracking beacon removed: A hidden <img> for tracking users was present in every raw article and has been stripped from content_html.
• See-also section stripped: Articles often ended with an "อ่านข่าว" (read more) section containing links to related articles. This section was removed from ~87 % of articles (1,049/1,200); the remaining ~13% either had no detectable header pattern or used edge-case formatting (e.g. <strong>-wrapped headers, inline typos).
• HTML elements retained in content_html: <p>, <h3>, <ol>, <ul>, <li>, <a>, <picture>, <img> (content images only), <iframe> (YouTube embeds), <blockquote>, <strong>, <span>, <br>.
• content_text is produced by calling BeautifulSoup's get_text(separator="") on the cleaned HTML, followed by whitespace normalisation: NO-BREAK SPACE (U+00A0) and EM SPACE (U+2003) are replaced with regular spaces, ZERO WIDTH SPACE (U+200B) and carriage returns are removed, and runs of consecutive spaces/tabs are collapsed.
• Stratified sampling for การเมือง: The 400 politics articles are drawn from three sub-groups — 40% (160) from the 2569 general-election cycle, 10% (40) from Cambodia border-conflict coverage, and 50% (200) general politics — to prevent two dominant news cycles from skewing the category distribution.

NLP Notes

• Thai script — no word boundaries: Thai does not use spaces between words. Space-split "token" counts are low (mean ~165 per article) and do not represent linguistic word counts. Use a Thai word segmenter such as PyThaiNLP (newmm or attacut) before any token-level processing.
• Date references: Articles frequently contain Thai-calendar year references (Buddhist Era, e.g. 2568 = 2025 CE). Add 543 to convert BE years to CE.
• Suitable tasks:
  • Text classification — three-class category prediction from title, abstract, or body
  • Summarization — abstract as a reference summary for content_text
  • Keyword / tag extraction — tags as a silver-standard label set
  • Thai language modelling — domain-specific news text spanning ~6 months

Source & License

Content is © Thai PBS. This dataset was collected from publicly accessible API endpoints for research and educational purposes. Users are responsible for complying with Thai PBS's terms of service before any redistribution or commercial use.