---
dataset_info:
  features:
  - name: words
    dtype: string
  - name: annotated_splits
    dtype: string
  - name: splits
    sequence:
      sequence: string
  - name: source
    dtype: string
  splits:
  - name: train
    num_bytes: 46318
    num_examples: 624
  download_size: 20677
  dataset_size: 46318
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
license: cc-by-4.0
language:
- da
pretty_name: Danish Compounds
size_categories:
- n<1K
---

## Danish Compounds

Danish Compounds consists of a list of Danish compound words annotated for constituent words and
interfixes. This can be used to evaluate how close a model can obtain similar splits.

## Loading the dataset

To load the dataset simply run:

```py
from datasets import load_dataset

ds = load_dataset("KennethEnevoldsen/danish-compounds", split="train")
```


## Sample

A sample from the dataset looks as follows:

```py
{
    "words": "A-aktie",
    "annotated_splits": "A|-|aktie",
    "splits": [["A"], ["-"], ["aktie"]],
    "source": "wiktionary",
}
```

Where:

- `words`: The word as a string
- `annotated_splits`: The annotated splits, including primary and secondary splits
- `splits`: The words.
- `source` Where the word was found.

## Dataset Source

The dataset is derived from two source [Wiktionary](https://da.wiktionary.org/wiki/Kategori:Sammensatte_ord_på_dansk) and [sproget.dk](https://sproget.dk/raad-og-regler/ordlister/sproglige-ordlister/ordforbindelser-i-et-eller-flere-ord).


## Annotation Creator

The word splits were manually annotated by Kenneth Enevoldsen.

## Annotation Scheme
Words that the annotator did not understand (e.g. some Icelandic words) were removed.

- "|" indicates primary split (into compound words)
- "&" indicates secondary split (used in case of compound word existing of a multiple compounds) e.g. blindtarmsbetændelse consist of two compounds "blindtarm" "betændelse" with a 'lim-s' ("s"), but blindtarm is itself a compound, namely "blind", "tarm".

## Reason for annotation

This dataset was initially annotated to get a gist of how well BPE, Unigram, and similar tokenizers obtain tokens similar to conventional splits.

![image/png](https://cdn-uploads.huggingface.co/production/uploads/5ff5943752c26e9bc240bada/tdTHo4W-_36-1U1gSR8CI.png)

> The Jaccard index, also sometimes called the Jaccard similarity, can be seen as a similarity metric
between the labeled splits and the splits of the model.

It was annotated as a part of Kenneth Enevoldsen's master thesis. 

## Citation

If you use this work please cite:

```
@mastersthesis{Enevoldsen2022,
  title     = {Representation Learning for Decision-Support in Clinical Psychiatry},
  author    = {Kenneth Enevoldsen},
  year      = {2022},
  school    = {Aarhus University},
  note      = {Master's Thesis}
}
```