Flair's German NER Model Hits 92.31 F1 With Document-Aware Context
flair/ner-german-large hits 92.31 F1 on revised CoNLL-03 German, 948K downloads. Predicts four tags: PER, LOC, ORG, MISC for German text. Built on XLM-RoBERTa large with document-level context via the FLERT approach. Minimal architecture: no CRF, no RNN, just one linear layer on top.
Takeaways
- flair/ner-german-large hits 92.31 F1 on revised CoNLL-03 German, 948K downloads.
- Predicts four tags: PER, LOC, ORG, MISC for German text.
- Built on XLM-RoBERTa large with document-level context via the FLERT approach.
- Minimal architecture: no CRF, no RNN, just one linear layer on top.
- Three-line usage via Flair, runs locally with no API costs attached.
- Ideal for GDPR redaction, knowledge graphs, and German news or legal extraction pipelines.
Flair’s German NER model extracts four entity types and reports an entity-level F1 score of 92.31 on the revised German CoNLL-03 benchmark. Its FLERT architecture can use neighboring sentences when classifying a span, giving developers a local model with structured output and a standard Python API.
Four labels, one large encoder
| Property | Value |
|---|---|
| Backbone | XLM-RoBERTa Large |
| Approximate parameters | 560 million |
| Entity labels | PER, LOC, ORG, MISC |
| Reported benchmark | 92.31 entity-level F1 |
| Training corpus | Revised German CoNLL-03 |
| Context | Neighboring sentences through FLERT |
| Output | Flat entity spans with labels and confidence scores |
The four labels cover people, locations, organizations, and miscellaneous named entities. CoNLL scoring requires an exact span boundary and the correct class, so a partially matched organization name receives no credit. The reported result provides an in-domain reference for newswire text; production accuracy depends on the target domain, preprocessing, and annotation policy.
Neighboring sentences supply extra evidence
Sentence-level NER can struggle when a name has several plausible meanings or appears in abbreviated form after an earlier mention. The FLERT paper addresses that limitation by placing surrounding document text into the transformer input and predicting labels for the target sentence.
Flair enables this behavior through use_context=True in its transformer embeddings. A standalone Sentence remains valid input and provides only its own tokens as evidence. Multi-sentence inference should preserve sentence order and document boundaries so the model can inspect related text without leaking context from another document.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.
Read original report
