Alibaba-NLP/KB-NER

Leverages a multilingual Wikipedia-based knowledge base with retrieval-augmented data augmentation and context enrichment to improve entity recognition across morphologically complex and code-mixed languages. Employs multi-stage fine-tuning with majority voting ensemble over transformer models (XLM-R) to integrate retrieved contextual information, with support for both paragraph and sentence-level knowledge retrieval strategies. Provides pre-processed datasets and trained models spanning 13 language tracks, enabling reproducible evaluation on the MultiCoNER benchmark.

186 stars. No commits in the last 6 months.