---
license: apache-2.0
tags:
- token-classification
- ner
- affiliation-parsing
- organization-extraction
- geoparsing
pipeline_tag: token-classification
base_model: xlm-roberta-base
widget:
- text: "MDPI, Grosspeteranlage 5, 4052 Basel, Switzerland."
- text: "Department of Chemistry, MIT, Cambridge, MA, USA."
---

# NER for affiliation parsing

`org-geo-ner` is a token-classification (NER) model that extracts organizations and geographic context from affiliation strings. It was fine-tuned in-house by [MDPI AG](https://www.mdpi.com/) to power [mdpi-ror-search](https://github.com/MDPI-AG/ror-search), an open-source tool that matches affiliation strings against the [Research Organization Registry (ROR)](https://ror.org).

## Model description

- **Architecture:** XLM-RoBERTa base (12 layers, 768 hidden size), fine-tuned for token classification
- **Task:** Named Entity Recognition (NER) over affiliation / address strings
- **Labeling scheme:** BIO (`B-`/`I-`/`O`)
- **Entity types:**
  - `ORG`: parent organization (e.g. university, company, ministry)
  - `SUB`: sub-organization / department, faculty, institute, lab
  - `CITY`: city
  - `COUNTRY`: country
- **Language:** primarily English, but trained on multilingual affiliation strings drawn from real scholarly manuscripts, so the model generalizes to non-English organization and place names to some extent
- **License:** Apache 2.0

## Intended use

This model is designed to extract structured entities from free-text author affiliation strings as found in scholarly manuscripts and metadata (e.g. `"Dept. of Physics, ETH, Zürich, Switzerland"`). It is the first stage in a pipeline whose downstream goal is to resolve affiliations to [ROR](https://ror.org) IDs. For further details about the pipeline, visit [mdpi-ror-search](https://github.com/MDPI-AG/ror-search).

It is **not** intended for general-purpose NER on arbitrary text (news, social media, etc.) as it is specialized for the affiliation/address domain.

## How to use

The model works out of the box with the `transformers` `token-classification` pipeline using `aggregation_strategy="simple"`, which merges wordpiece tokens into full entity spans.

```python
from transformers import pipeline

ner_pipeline = pipeline(
    "token-classification",
    model="mdpi-ai/org-geo-ner",
    aggregation_strategy="simple",
)

text = "MDPI, Grosspeteranlage 5, 4052 Basel, Switzerland."
output = ner_pipeline(text)

for entity in output:
    print(f"{entity['entity_group']:<10} {entity['word']:<40} score={entity['score']:.3f}")
```

Output:

```
ORG             MDPI                           score=0.855
CITY            Basel                          score=0.994
COUNTRY         Swtizerland                    score=1.000
```

Batch inference works the same way by passing a list of strings:

```python
texts = [
    "MDPI, Grosspeteranlage 5, 4052 Basel, Switzerland.",
    "Department of Chemistry, MIT, Cambridge, MA, USA.",
]
outputs = ner_pipeline(texts)
```

For production use, filter low-confidence predictions with a score threshold (e.g. `score >= 0.65`) before downstream processing.

## Training data

The model was fine-tuned on an **internal, proprietary dataset** of affiliation strings annotated for `ORG`, `SUB`, `CITY`, and `COUNTRY` spans. This dataset is not publicly released.

## Evaluation

Evaluated on a held-out internal test split, per-entity metrics (strict span matching):

| Entity   | Precision | Recall | F1    |
|----------|-----------|--------|-------|
| ORG      | 94.7      | 95.4   | 95.4  |
| SUB      | 94.2      | 95.9   | 95.0  |
| CITY     | 97.2      | 97.4   | 97.3  |
| COUNTRY  | 98.2      | 99.1   | 98.7  |
| **Overall** | **95.9** | **96.8** | **96.4** |

## Limitations and bias

- Trained on affiliation strings sourced from scholarly manuscripts; performance may degrade on out-of-domain text.
- The training data, while multilingual in the organization and place names it contains, is predominantly structured around English-language affiliation conventions; performance on affiliations written entirely in non-Latin scripts is not guaranteed.
- The training dataset is internal and not publicly available, so independent reproduction of the reported metrics is not possible.
- As with any NER model, ambiguous or heavily abbreviated affiliation strings (e.g. bare acronyms) may be mislabeled or missed; downstream consumers should apply a confidence threshold and treat predictions as candidates rather than ground truth.
- With `aggregation_strategy="simple"`, hyphenated compound organization names (e.g. `CREAF-CSIC-UAB`) can be split into multiple adjacent `ORG` spans rather than merged into one, since the underlying XLM-RoBERTa tokenizer treats `-` as a token boundary. Downstream consumers may want to re-merge adjacent same-label spans separated only by punctuation.

## Additional information

- **Repository (pipeline code, docs):** [MDPI-ROR-Search](https://github.com/MDPI-AG/ror-search).
- **Contact:** For questions, issues, or collaboration inquiries, please open an issue on the [GitHub repository](https://github.com/MDPI-AG/ror-search).
- **Funding:** This project was fully funded and supported by [MDPI AG](https://www.mdpi.com/).
- **License:** [Apache 2.0](https://www.apache.org/licenses/LICENSE-2.0)
- **Citation:** No associated publication is available at this time. If you use this model, please cite it as:

```bibtex
@misc{orggeoner,
  title  = {org-geo-ner: A Named Entity Recognition Model for Affiliation Parsing},
  author = {MDPI AG},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/mdpi-ai/org-geo-ner}}
}
```

## Disclaimer

This model is provided as-is, without warranty of any kind. Predictions should be reviewed before being used in contexts requiring high accuracy, such as automated compliance or reporting workflows.