
Korea Press Foundation
KPF-BERT and AI-based News Services
We led the development of KPF-BERT, a language model specialised for the news domain, and built the news analysis service and the large language model pilot service that use it.
Jul 2021 – Dec 2025

Project
- Jul–Dec 2021 — language resource development for news organisations.
- Feb 2022 – May 2024 — maintenance and development of the NewsAlgo service.
- Sep–Nov 2022 — a temporary site for the forum on improving digital news distribution.
- Aug–Oct 2023 — a pilot service built on a large language model.
- 2025 — collaboration on collecting and analysing YouTube journalism data.
Challenge
- News articles are unstructured text in bulk, so they must first be put into a form a machine can work with.
- A general-purpose language model does not capture the vocabulary and register specific to news writing.
- Looking only at the output of a news ranking algorithm does not reveal the criteria it applied.
Development
Language model and NLP
- Led the development of KPF-BERT, trained on twenty years of news articles held by the Korea Press Foundation.
- Article body summarisation.
- Article clustering.
- Text-based natural language processing solutions.
Service development
- NewsAlgo — a media literacy site where students see for themselves how a news ranking algorithm orders articles. We handled its maintenance and development.
- The LLM pilot service — automatically extracts figures from an article body and converts them into tables and charts.
- The site used for the forum on improving digital news distribution.
Data analysis collaboration
- Selected the top 200 YouTube channels in news and politics.
- Collected and pre-processed the comment data.
- Ran sentiment analysis and keyword extraction.
- Delivered the viewer response patterns as evidence for the forum.
Screens


These are the live product screens. The interface is in Korean.