Rishabh157/unified-multilingual-ner-mdeberta

Unified Multilingual NER (mDeBERTa-v3-base)

This model is a fine-grained Multilingual Named Entity Recognition (NER) model fine-tuned on 1.53+ million sentences across three major benchmark suites:

• WikiANN
• MultiCoNER 2023 (v2)
• MultiNERD

Built on top of microsoft/mdeberta-v3-base, it supports 47 fine-grained entity categories (95 BIO logit tags) across 12+ languages.

Developed by Rishabh Kumar.

---

1. Model Description & Architecture

• Base Model: microsoft/mdeberta-v3-base (278M parameters, 12 Transformer layers, 768 hidden size, 250k vocabulary)
• Fine-Tuning Method: LoRA (Low-Rank Adaptation, r=64, α=64, target modules: query_proj, key_proj, value_proj, intermediate.dense, output.dense) merged into standalone base weights.
• Precision: FP16 (torch.float16, ~530 MB footprint)
• Task: Multilingual Token Classification / Named Entity Recognition
• Languages Supported (15): English (en), Spanish (es), French (fr), German (de), Italian (it), Portuguese (pt), Swedish (sv), Dutch (nl), Polish (pl), Russian (ru), Farsi (fa), Bengali (bn), Hindi (hi), Ukrainian (uk), Chinese (zh), and multilingual mixed text.

---

2. Dataset Links & Entity Taxonomy

The model combines the taxonomy of three major Hugging Face NER datasets into a single unified classification head (47 × 2 + 1 = 95 BIO output logits):

A. WikiANN (3 Core Broad Categories)

• Languages Included (7): German (de), English (en), Spanish (es), French (fr), Italian (it), Portuguese (pt), Swedish (sv)
• Entity Classes: PER (Person), LOC (Location), ORG (Organization)

B. MultiNERD (15 Multi-Domain Categories)

• Languages Included (10): German (de), English (en), Spanish (es), French (fr), Italian (it), Dutch (nl), Polish (pl), Portuguese (pt), Russian (ru), Chinese (zh)
• Standard: PER, LOC, ORG, TIME
• Specialized Domains:
  • ANIM (Animal species, breeds, fauna)
  • PLANT (Plants, trees, botanical species, fungi)
  • FOOD (Food, dishes, beverages, culinary items)
  • DIS (Diseases, disorders, medical conditions)
  • CEL (Celestial bodies, stars, planets, galaxies)
  • EVE (Historical, sports, cultural events)
  • INST (Public institutions, international bodies)
  • MEDIA (Newspapers, networks, press outlets)
  • MYTH (Mythological entities, deities, folklore)
  • VEHI (Vehicles, ships, aircraft, spacecraft)
  • BIO (Biological compounds, genes, proteins)

C. MultiCoNER 2023 (v2) (33 Fine-Grained Categories)

• Languages Included (13): Bangla (BN), German (DE), English (EN), Spanish (ES), Farsi (FA), French (FR), Hindi (HI), Italian (IT), Portuguese (PT), Swedish (SV), Ukrainian (UK), Chinese (ZH), and Multilingual Mixed (MULTI)
• Person Subtypes: Artist, Athlete, Cleric, Politician, Scientist, SportsManager, OtherPER
• Organization Subtypes: AerospaceManufacturer, CarManufacturer, MusicalGRP, PrivateCorp, PublicCorp, SportsGRP
• Location Subtypes: Facility, HumanSettlement, Station, OtherLOC
• Products & Creative Works: ArtWork, MusicalWork, VisualWork, WrittenWork, Software, Clothing, Drink, Food, OtherPROD
• Medical & Healthcare: AnatomicalStructure, Disease, MedicalProcedure, Medication/Vaccine, Symptom

D. Complete Unified 47 Entity Category List

 1. ANIM                   13. Disease               25. Medication/Vaccine    37. Scientist
 2. AerospaceManufacturer  14. Drink                 26. MusicalGRP            38. Software
 3. AnatomicalStructure    15. EVE                   27. MusicalWork           39. SportsGRP
 4. ArtWork                16. FOOD                  28. ORG                   40. SportsManager
 5. Artist                 17. Facility              29. OtherLOC              41. Station
 6. Athlete                18. Food                  30. OtherPER              42. Symptom
 7. BIO                    19. HumanSettlement       31. OtherPROD             43. TIME
 8. CEL                    20. INST                  32. PER                   44. VEHI
 9. CarManufacturer        21. Cleric                33. PLANT                 45. Vehicle
10. Cleric                 22. MEDIA                 34. Politician            46. VisualWork
11. Clothing               23. MYTH                  35. PrivateCorp           47. WrittenWork
12. DIS                    24. MedicalProcedure      36. PublicCorp

---

3. Training Results & Metrics

The model was fine-tuned for 5 epochs (~28.3 GPU hours on an NVIDIA L4 GPU) using AdamW (lr=2e-5, linear scheduler, effective batch size 16).

Final Validation Evaluation (Epoch 5 / 5)

Metric	Score
Validation Entity F1	0.7693
Validation Precision	0.8449
Validation Recall	0.7061
Validation Accuracy	0.9587
Validation Loss	0.0276

Per-Epoch Progression

• Epoch 1: F1 = 0.7183 | Precision = 0.8016 | Recall = 0.6506 | Loss = 0.0372
• Epoch 2: F1 = 0.7322 | Precision = 0.8151 | Recall = 0.6645 | Loss = 0.0317
• Epoch 3: F1 = 0.7544 | Precision = 0.8346 | Recall = 0.6883 | Loss = 0.0293
• Epoch 4: F1 = 0.7669 | Precision = 0.8460 | Recall = 0.7013 | Loss = 0.0281
• Epoch 5: F1 = 0.7693 | Precision = 0.8449 | Recall = 0.7061 | Loss = 0.0276

Independent Test Benchmarks (32-Suite)

Evaluation across WikiANN, MultiNERD (10 languages + full), and MultiCoNER 2023 (13 languages + full):

Dataset	Subset / Lang	Sentences	Precision	Recall	F1 Score	Accuracy
combined	combined	536,661	0.6639	0.6298	0.6464	0.9313
multinerd	🔹 multinerd_full	167,993	0.8906	0.7298	0.8022	0.9646
multinerd	🇬🇧 EN-English	16,454	0.9097	0.8103	0.8571	0.9771
multinerd	🇩🇪 DE-German	15,762	0.9025	0.7874	0.8410	0.9761
multinerd	🇷🇺 RU-Russian	8,338	0.8211	0.7170	0.7656	0.9631
multinerd	🇳🇱 NL-Dutch	17,181	0.8313	0.6641	0.7384	0.9647
multinerd	🇫🇷 FR-French	17,695	0.7738	0.6725	0.7196	0.9668
multinerd	🇵🇱 PL-Polish	19,555	0.8046	0.6369	0.7110	0.9510
multinerd	🇪🇸 ES-Spanish	17,399	0.7893	0.6374	0.7053	0.9653
multinerd	🇵🇹 PT-Portuguese	17,815	0.7471	0.6083	0.6706	0.9585
multinerd	🇮🇹 IT-Italian	18,217	0.7424	0.5844	0.6540	0.9619
multinerd	🇨🇳 ZH-Chinese	19,577	0.3193	0.2353	0.2709	0.8184
wikiann	🔹 wikiann_full	70,000	0.6214	0.6199	0.6202	0.8099
wikiann	sv (Swedish)	10,000	0.7174	0.6662	0.6909	0.8843
wikiann	de (German)	10,000	0.6874	0.6352	0.6603	0.8938
wikiann	it (Italian)	10,000	0.6506	0.6459	0.6482	0.8362
wikiann	es (Spanish)	10,000	0.6076	0.6560	0.6309	0.7726
wikiann	fr (French)	10,000	0.6017	0.6080	0.6049	0.7581
wikiann	pt (Portuguese)	10,000	0.5613	0.5976	0.5789	0.7517
wikiann	en (English)	10,000	0.5238	0.5304	0.5271	0.7729
multiconer2023	🔹 multiconer2023_full	1,991,359	0.5957	0.5923	0.5937	0.9160
multiconer2023	IT-Italian	247,881	0.6566	0.6654	0.6609	0.9225
multiconer2023	SV-Swedish	231,190	0.6468	0.6332	0.6400	0.9182
multiconer2023	UK-Ukrainian	238,296	0.6123	0.5994	0.6058	0.9232
multiconer2023	FR-French	249,786	0.6039	0.5971	0.6005	0.9122
multiconer2023	ES-Spanish	246,900	0.5889	0.6060	0.5974	0.9212
multiconer2023	PT-Portuguese	229,490	0.5928	0.5987	0.5957	0.9197
multiconer2023	MULTI-Multilingual	358,668	0.5787	0.5840	0.5813	0.9118
multiconer2023	DE-German	20,145	0.5625	0.5993	0.5803	0.9148
multiconer2023	EN-English	249,980	0.5620	0.5818	0.5718	0.9078
multiconer2023	HI-Hindi	18,399	0.5308	0.5575	0.5438	0.9312
multiconer2023	BN-Bangla	19,859	0.5029	0.5275	0.5149	0.9079
multiconer2023	FA-Farsi	219,168	0.5286	0.4606	0.4923	0.9072
multiconer2023	ZH-Chinese	20,265	0.4129	0.4904	0.4483	0.8619

---

4. Inference Mechanism & Softmax Logit Decoding (95 Logit Classes)

During inference, for any input text, our fine-tuned model (microsoft/mdeberta-v3-base with merged LoRA parameters) passes subword token representations through its 12 Transformer layers to output a 95-dimensional logit vector per token. PyTorch applies a softmax function across all 95 output tags to compute normalized probability confidence scores:

P(tag_k | token_i) = exp(z_k) / Σ exp(z_j)   for k ∈ {1, 2, ..., 95}

• Hard Argmax (Default Token Classification): Selects the single highest-probability tag (argmax).
• Soft Confidence Distribution: For ambiguous entity tokens (e.g. "Harvard"), the model distributes probability across related candidate classes (B-INST: 72.4%, B-ORG: 21.1%, B-Facility: 5.3%), capturing semantic cross-taxonomy relationships.

---

5. Intended Use & Out-of-Scope Use

Intended Use

• Information extraction from multilingual text (news articles, Wikipedia, social media, web content)
• Downstream NLP pipelines requiring structured entity extraction across multiple languages
• Research in cross-lingual transfer learning, fine-grained NER, and multilingual NLP
• Knowledge graph construction from unstructured multilingual text

Out-of-Scope Use

• Domain-specific NER (e.g., legal, financial, or clinical NER) — may require additional fine-tuning on domain data
• Languages not listed in the supported 15-language set — the model was not trained on those languages

---

6. How to Use

Quick Inference via Hugging Face Pipeline

from transformers import pipeline

ner_pipeline = pipeline(
    "token-classification",
    model="Rishabh157/unified-multilingual-ner-mdeberta",
    aggregation_strategy="simple"
)

text = "Barack Obama visited the United Nations headquarters in New York City."
entities = ner_pipeline(text)

for entity in entities:
    print(f"{entity['word']:<25} | {entity['entity_group']:<15} | Score: {entity['score']:.4f}")

PyTorch & Transformers Direct Usage

import torch
from transformers import AutoTokenizer, AutoModelForTokenClassification

model_name = "Rishabh157/unified-multilingual-ner-mdeberta"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)

text = "Emmanuel Macron est le président de la République française à Paris."
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.argmax(outputs.logits, dim=2)

tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])
labels = [model.config.id2label[p.item()] for p in predictions[0]]

for token, label in zip(tokens, labels):
    if label != "O":
        print(f"{token:<20} -> {label}")

---

7. Known Considerations & Future Improvements

• European Languages Lead Performance: The model achieves highest F1 scores on European languages (English, German, Dutch, Russian — F1 0.76–0.86 on MultiNERD) due to richer training data availability.
• East Asian & South Asian Languages: Chinese (ZH), Farsi (FA), Bangla (BN), and Hindi (HI) show room for improvement and can benefit from additional fine-tuning on language-specific corpora.
• Unified Taxonomy Design: The model unifies 47 entity types from three separate datasets into a single classification head. Some categories across datasets are semantically related (e.g., Food from MultiCoNER and FOOD from MultiNERD) and are treated as distinct classes to preserve dataset fidelity.
• Continual Improvement: Performance on lower-resource languages can be improved by fine-tuning further on domain-specific or language-specific NER datasets.

---

8. Citation

If you use this model or dataset pipeline in your research, please cite:

@misc{kumar2026multilingualner,
  author = {Rishabh Kumar},
  title = {Unified Fine-Grained Multilingual NER with mDeBERTa-v3},
  year = {2026},
  publisher = {Hugging Face},
  howpublished = {\url{https://huggingface.co/Rishabh157/unified-multilingual-ner-mdeberta}}
}