K-BERT

Why we included this project

K-BERT is the reference implementation of the AAAI 2020 paper that taught a BERT-style model to read knowledge graphs alongside text. It wraps triples from sources like CN-DBpedia and HowNet into extra tokens, so during fine-tuning the model sees factual background next to the original sentence. The repo ships two runnable scripts, one for text classification and one for named entity recognition, with links to the pretrained weights and knowledge files. Since it sits on the UER framework, the code stays compact and is a decent starting point for anyone experimenting with knowledge-aware fine-tuning, especially for Chinese-language tasks.