README.md

---
annotations_creators:
- expert-generated
- crowdsourced
language:
- nl
language_creators:
- machine-generated
license:
- cc-by-2.0
multilinguality:
- monolingual
pretty_name: Contentious Contexts Corpus
size_categories:
- 1K<n<10K
source_datasets:
- original
tags:
- newspapers
- historic
- dutch
- problematic
- ConConCor
task_categories:
- text-classification
task_ids:
- sentiment-scoring
- multi-label-classification
dataset_info:
  features:
  - name: extract_id
    dtype: string
  - name: text
    dtype: string
  - name: target
    dtype: string
  - name: annotator_responses_english
    list:
    - name: id
      dtype: string
    - name: response
      dtype: string
  - name: annotator_responses_dutch
    list:
    - name: id
      dtype: string
    - name: response
      dtype: string
  - name: annotator_suggestions
    list:
    - name: id
      dtype: string
    - name: suggestion
      dtype: string
  splits:
  - name: train
    num_bytes: 2199414
    num_examples: 2720
  download_size: 1940545
  dataset_size: 2199414
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
---

# Contentious Contexts Corpus (ConConCor)

2,720 short extracts from Dutch newspapers published between 1890 and 1940, each built around a target word that may read as contentious today, with 21,800 annotations from 436 annotators judging whether that word's use in that context is contentious by present-day standards.

Built by [Cultural AI](https://github.com/cultural-ai) — a collaboration of the KNAW Humanities Cluster and Centrum Wiskunde & Informatica — for the EuropeanaTech Challenge for Europeana AI/ML datasets, and described in Brate et al., [Capturing Contentiousness](https://doi.org/10.1145/3460210.3493553) (K-CAP 2021). The creators' own datasheet is [in the source repo](https://github.com/cultural-ai/ConConCor/blob/main/Dataset/DataSheet.pdf).

The motivating problem is a curatorial one. Archives, libraries and museums hold descriptions and texts written in language that is offensive now and was unremarkable then, and they mostly want to keep the original wording while flagging it rather than silently rewriting it. Doing that at collection scale needs a way to tell which uses are contentious, which is what this corpus is for.

## What is in the extracts

The extracts are OCR'd newspaper articles and they contain, in ordinary reportage, the vocabulary of Dutch colonial and early-20th-century society: racial and ethnic slurs, colonial terms for people in the Dutch East Indies and Suriname, terms for disability, and period terms for sexuality and religion. There are 91 target words in all. Some are neutral by design — `achtergrond`, `historisch`, `westers`, `etniciteit` — and colonial and current place names are paired (`batavia`/`jakarta`), so the corpus is not a list of slurs with sentences attached.

The annotations are judgements, not facts. Annotators were asked whether a use is contentious *from today's perspective*; they disagree, and the disagreement is the signal the corpus was built to record.

## Structure

One `train` split of 2,720 rows.

| field | contents |
|---|---|
| `extract_id` | Identifier for the extract |
| `text` | The extract, roughly five sentences of OCR'd article text |
| `target` | The target word as it appears in `text` |
| `annotator_responses_english` | List of `{id, response}`, one per annotator |
| `annotator_responses_dutch` | The same responses in the original Dutch |
| `annotator_suggestions` | List of `{id, suggestion}` — free-text alternatives or other contentious terms spotted, Dutch only, often empty |

The four responses, and how often each was given across all 21,800 annotations:

| Dutch | English | count |
|---|---|---|
| Niet omstreden | Not contentious | 13,386 (61.4%) |
| Omstreden naar huidige maatstaven | Contentious according to current standards | 4,804 (22.0%) |
| Weet ik niet | I don't know | 2,210 (10.1%) |
| Onleesbare OCR | Illegible OCR | 1,400 (6.4%) |

## Things to check before you use it

**There is no gold label, deliberately.** Aggregation is left to you, because how you aggregate is a research choice. A plain majority vote over contentious-vs-not gives 496 extracts contentious, 2,070 not, and 154 tied.

**The target word is marked with Unicode mathematical bold italic characters** (U+1D400 block), not markup: `𝙄𝙣𝙙𝙤-𝙀𝙪𝙧𝙤𝙥𝙚𝙖𝙖𝙣`, not `Indo-Europeaan`. Tokenisers will fragment those, and matching `target` against `text` fails until you normalise. `unicodedata.normalize("NFKC", text)` maps them back to ASCII — do it before anything else, and note that it also erases the marking, so record the offsets first if you need them.

**Five extracts have 416 annotations each; the median is 7.** `c0` to `c4` were shown to every annotator as shared calibration items. They will dominate any annotator-agreement calculation that treats extracts as equally weighted. At the other end, 45 extracts carry a single annotation.

**6.4% of annotations say the OCR is illegible.** This is Europeana newspaper OCR from 1890–1940 originals, and quality varies. Counting `Illegible OCR` votes per extract is a usable proxy for OCR quality, and a reasonable filter.

**Expert and crowd annotators are not distinguished in these fields.** The paper's central finding is that experts and crowdworkers judge contentiousness differently, but the HF fields carry only an anonymised ID. In the source repo the 413 Prolific crowdworkers are those listed in `Demographics.csv`; the remaining 23 IDs are not. Reconstructing the split means going back to the source CSVs.

**It is Dutch, and it is dated.** Judgements were collected in 2021 from present-day Dutch speakers about 1890–1940 Dutch text. Neither the vocabulary nor the sensibilities transfer to other languages, other periods, or a moderation task on contemporary text.

## Aggregating the annotations

```python
from collections import Counter

def ocr_score(example):
    """1.0 = no annotator found the OCR illegible."""
    responses = [r["response"] for r in example["annotator_responses_english"]]
    bad = Counter(responses).get("Illegible OCR", 0)
    return round(1 - bad / len(responses), 3)

def majority_vote(example):
    """Returns 'contentious', 'not_contentious', or 'tied'."""
    counts = Counter(r["response"] for r in example["annotator_responses_english"])
    c = counts.get("Contentious according to current standards", 0)
    n = counts.get("Not contentious", 0)
    return "contentious" if c > n else "not_contentious" if n > c else "tied"
```

## Loading

```python
from datasets import load_dataset

ds = load_dataset("biglam/contentious_contexts", split="train")
row = ds[0]
row["text"]                          # the newspaper extract
row["target"]                        # the flagged compound, bolded in `text`
row["annotator_responses_dutch"]     # [{'id': ..., 'response': 'Niet omstreden'}, ...]
row["annotator_responses_english"]   # the same responses, translated
```

`Metadata.csv` in the [source repo](https://github.com/cultural-ai/ConConCor) adds newspaper
title, date and place for 2,719 of the 2,720 extracts; it is not carried here.

## Licence

CC-BY. The source newspaper images and OCR come from the Europeana Newspaper collection as provided by the KB, National Library of the Netherlands.

## Citation

```bibtex
@inproceedings{brate2021capturing,
  author    = {Brate, Ryan and Nesterov, Andrei and Vogelmann, Valentin and
               van Ossenbruggen, Jacco and Hollink, Laura and van Erp, Marieke},
  title     = {Capturing Contentiousness: Constructing the Contentious Terms in Context Corpus},
  booktitle = {Proceedings of the 11th on Knowledge Capture Conference (K-CAP '21)},
  year      = {2021},
  pages     = {17--24},
  doi       = {10.1145/3460210.3493553}
}

@misc{ContentiousContextsCorpus2021,
  author       = {Cultural AI},
  title        = {Contentious Contexts Corpus},
  year         = {2021},
  publisher    = {GitHub},
  howpublished = {\url{https://github.com/cultural-ai/ConConCor}},
}
```