deepakint/knowledge-platform-ner

Knowledge Platform NER

A cross-domain, multilingual Named Entity Recognition model built for the Knowledge Platform — a system that connects patents, scientific papers, news articles, and political documents across 13 data sources.

Fine-tuned from answerdotai/ModernBERT-base on 256K+ multilingual documents spanning patents (USPTO, EPO), scientific papers (OpenAlex, arXiv), political documents (Bundestag, EU Parliament), and news.

Key Results

Metric	Score
F1	90.6%
Precision	89.5%
Recall	91.8%
Accuracy	98.1%

Entity Types

The model recognizes 15 entity types using BIO tagging (31 labels total):

Tag	Entity Type	Example
PER	Person	James Chen, Lisa Paus, Yann LeCun
ORG	Organization	Samsung Electronics, Bundestag, OpenAI
LOC	Location	Seoul, Brüssel, New York
ANIM	Animal	E. coli, SARS-CoV-2
BIO	Biological	CRISPR-Cas9, mRNA
CEL	Celestial Body	Mars, Jupiter
DIS	Disease	Alzheimer's, sickle cell disease
EVE	Event	COP28, World Economic Forum
FOOD	Food	glyphosate, insulin
INST	Instrument	LiDAR, mass spectrometer
MEDIA	Media/Work	Nature, The Lancet
MYTH	Mythological	Apollo (program context)
PLANT	Plant	Arabidopsis, cannabis sativa
TIME	Time	Q3 2025, fiscal year 2024
VEHI	Vehicle	Falcon 9, Boeing 787

Use Cases

This model is designed for knowledge graph construction from heterogeneous document collections:

• Patent Analysis: Extract assignees, inventors, locations, and technologies from patent filings
• Scientific Literature: Identify authors, institutions, biological entities, and instruments from papers
• Political Document Processing: Extract politicians, parties, organizations from parliamentary debates (EN + DE)
• News Processing: Identify key entities across news articles for event tracking
• Cross-Domain Knowledge Graphs: Connect entities that appear across different document types and languages

Works with the Knowledge Platform Embedding Model

This model is designed to work alongside deepakint/knowledge-platform-embeddings — a SciNCL-based embedding model fine-tuned with contrastive learning on the same document corpus.

Together they form a pipeline:

1. This NER model extracts entities (the nodes of a knowledge graph)
2. The embedding model finds document connections (the edges of a knowledge graph)

Quick Start

from transformers import pipeline

ner = pipeline(
    "ner",
    model="deepakint/knowledge-platform-ner",
    aggregation_strategy="max"
)

# English patent text
text = "Samsung Electronics Co., Ltd. filed a patent at the USPTO in Washington, D.C."
entities = ner(text)

for entity in entities:
    print(f"  {entity['word']:40s} {entity['entity_group']:10s} {entity['score']:.3f}")

  Samsung Electronics Co., Ltd.            ORG        1.000
  USPTO                                    ORG        0.998
  Washington, D.C.                         LOC        0.999

# German political text
text = "Lisa Paus sprach im Deutschen Bundestag in Berlin über die neue Regulierung."
entities = ner(text)

for entity in entities:
    print(f"  {entity['word']:40s} {entity['entity_group']:10s} {entity['score']:.3f}")

  Lisa Paus                                PER        1.000
  Deutschen Bundestag                      ORG        1.000
  Berlin                                   LOC        1.000

Grouping Entities by Type

from collections import defaultdict

text = """Apple Inc. CEO Tim Cook announced a new research lab in Palo Alto, 
California, partnering with Stanford University on CRISPR gene editing research."""

entities = ner(text)
grouped = defaultdict(list)
for ent in entities:
    grouped[ent["entity_group"]].append(ent["word"])

for label, names in sorted(grouped.items()):
    print(f"  {label:8s}: {names}")

  BIO     : ['CRISPR']
  LOC     : ['Palo Alto', 'California']
  ORG     : ['Apple Inc.', 'Stanford University']
  PER     : ['Tim Cook']

Training Details

Base Model

answerdotai/ModernBERT-base — a 149M parameter encoder model with:

• 8,192 token context length (vs. 512 for classic BERT)
• Rotary Position Embeddings (RoPE)
• Alternating full + sliding window attention
• Pre-trained on 2 trillion tokens of English text

Training Data

~256,000 documents from 13 data sources across multiple domains and languages:

Domain	Sources	Language
Patents	USPTO, EPO	EN, DE
Scientific Papers	OpenAlex, arXiv	EN
Political Documents	Bundestag, EU Parliament	DE, EN
News	Various	EN, DE

Hyperparameters

Parameter	Value
Learning rate	2e-05
Batch size	16 (x2 gradient accumulation = 32 effective)
Epochs	3
Optimizer	AdamW
LR scheduler	Cosine with 10% warmup
Seed	42

Training Progress

Epoch	Training Loss	Validation Loss	Precision	Recall	F1	Accuracy
1	0.1276	0.0766	0.8595	0.8361	0.8476	0.9728
2	0.0927	0.0623	0.8659	0.8923	0.8789	0.9777
3	0.0422	0.0694	0.8707	0.8949	0.8827	0.9778

Note: The best checkpoint (epoch ~2, lowest validation loss 0.0606) was selected as the final model, achieving 90.6% F1.

Strengths and Limitations

Strengths

• Cross-domain: Works on patents, papers, news, and political documents with a single model
• Multilingual: Handles both English and German text
• Rich entity types: 15 entity types covering people, organizations, locations, biological entities, diseases, instruments, and more
• Fast: ~5ms per document on CPU — suitable for processing millions of documents
• Long context: Inherits ModernBERT's 8,192 token context window

Limitations

• Conference/product names: May fragment uncommon compound names (e.g., "NeurIPS" split into tokens) — use confidence thresholding (>0.5) to filter
• Languages: Optimized for English and German; other languages may work but are untested
• Domain drift: Performance is best on patent, scientific, political, and news text — may degrade on informal text (social media, chat)

Recommended Post-Processing

For production use, apply a confidence threshold to filter low-quality predictions:

# Filter entities with confidence > 0.5
entities = [e for e in ner(text) if e["score"] > 0.5]

Framework Versions

• Transformers: 5.6.0
• PyTorch: 2.5.1+cu121
• Datasets: 4.8.4
• Tokenizers: 0.22.2

Citation

@misc{knowledge-platform-ner-2026,
  title={Knowledge Platform NER: Cross-Domain Multilingual Named Entity Recognition},
  author={deepakint},
  year={2026},
  url={https://huggingface.co/deepakint/knowledge-platform-ner}
}

Related Models

• Embedding Model: deepakint/knowledge-platform-embeddings — Cross-domain semantic search and document matching