README.md

---
license: mit
task_categories:
- token-classification
language:
- nl
tags:
- digital_humanities
size_categories:
- 1K<n<10K
---
# Dataset Card for Dataset Name

The globalise_NER_token_classification dataset is a fine-grained dataset for the training of token-classification NER models on Dutch East-India Company texts (17th to 18th century).

## Dataset Details

### Dataset Description

The dataset provides 15 fine-grained labels detailing activities and people of the Dutch East-India Company (VOC), and can be used to train NER token-classification models for the 
period 17th-18th century and the domain of VOC texts. The texts are taken from the [Overgebleven Brieven & Papieren](https://globalise.huygens.knaw.nl/source-corpus/) corpus, 
and preprocessed for annotation as described in [(Arnoult et al., 2025)](#citation).

- **Curated by:** Brecht Nijman
- **Funded by:** Dutch Research Council (NWO)
- **Shared by:** Globalise team
- **Language(s) (NLP):** nl (Early Modern Dutch)
- **License:** MIT

### Dataset Sources

- **Repository:** [globalise-huygens/finegrained-hist-ner](https://github.com/globalise-huygens/finegrained-hist-ner)
- **Paper:** [tbd]

## Uses

The dataset is intended for training NER token-classification models for Early Modern Dutch in the VOC domain.

### Direct Use

Training or data-augmentation for Dutch historical NER models.

### Out-of-Scope Use

The training data represents a historical variant of Dutch and a restricted domain (VOC documents), and is not expected to be useful for other variants of Dutch or domains.

## Dataset Structure

Annotations were collected in several rounds: a first part of the training data was collected together with the validation data, with a random split based on token sequences;
 a later round of annotations was split by document between additional training data and test data. See [(Arnoult et al., 2025)](#citation) for more details.

| | train | validation | test |
| --- | --- | --- | --- |
| sequences | 576 | 78 | 98 |
| tokens | 44695 | 6001 | 10133 |
| entities | 5932 | 893 | 887 |


## Dataset Creation

### Curation Rationale

The dataset was created to provide domain-specific NER labels for the processing of the Overgebleven Brieven & Papieren corpus.

### Source Data

The [Overgebleven Brieven & Papieren](https://globalise.huygens.knaw.nl/source-corpus/) corpus, a collection of various documents written under the VOC administration: reports, 
ship inventories, letters, etc.

#### Data Collection and Processing

The source corpus was preprocessed as follows:

* HTR: with Loghi ([Koert et al. 2024](https://doi.org/10.1007/978-3-031-70645-5_6))
* word tokenization: with SpaCy [nl_core_news_lg](https://spacy.io/models/nl/#nl_core_news_lg)

Documents were manually reconstructed from page scans to provide coherent contexts for annotations. 26 documents were selected for annotation, spanning the years 1618 to 1782.

#### Who are the source data producers?

The corpus texts were written by the VOC administration, and later collected by the [Huygens Instituut](https://www.huygens.knaw.nl/en/) and its predecessors. 

### Annotations

The annotations enrich the texts with 15 NER tags, identifying common entity types (persons, locations, organisations), VOC-domain types (commodities, ships) and providing 
fine-grained types for people (profession, status). The tagset is described further in [(Arnoult et al., 2025)](#citation).

#### Annotation process

See [(Arnoult et al., 2025)](#citation).

#### Who are the annotators?

idem.

#### Personal and Sensitive Information

The dataset contains personal information about past people.

## Bias, Risks, and Limitations

The source corpus is biased, representing the standpoint of an early colonial organisation that notably used military force and engaged in slave trade.

### Recommendations

Users should be aware of the often violent character of the texts underlying the annotations.

## Citation

**BibTeX:**

[tbd]

**APA:**

[tbd]


## Dataset Card Contact

Sophie Arnoult