metadata

task_categories:
  - text-generation
  - fill-mask
language:
  - 'no'
  - nb
  - nn
pretty_name: Norwegian Idioms
size_categories:
  - 1K<n<10K
license: cc0-1.0
configs:
  - config_name: default
    data_files:
      - split: test
        path: data.jsonl
  - config_name: include_translated_idioms
    data_files:
      - split: test
        path: original_and_translated_data.jsonl

NorEval: NorIdiom

This dataset is a part of the NorEval evaluation suite.
See the NorEval codebase here: https://github.com/ltgoslo/noreval
Read the preprint here: https://arxiv.org/abs/2504.07749

@article{mikhailov2025noreval,
  title={NorEval: A Norwegian Language Understanding and Generation Evaluation Benchmark},
  author={Mikhailov, Vladislav and Enstad, Tita and Samuel, David and Farseth{\aa}s, Hans Christian and Kutuzov, Andrey and Velldal, Erik and {\O}vrelid, Lilja},
  journal={arXiv preprint arXiv:2504.07749},
  year={2025}
}

Norwegian idioms

This is a dataset that consists of 3553 Norwegian idioms and phrases that appear more than 100 times in the online library of the National Library of Norway.
There are 3472 Bokmål idioms and 89 Nynorsk idioms.

To improve the Bokmål/Nynorsk imbalance, we have automatically translated all idioms that exist in one language and not in the other with apertium. These have again been filtered on frequency in the online library. Including the translated idioms, there are 4600 unique idioms; 3502 Bokmål idioms and 1707 Nynorsk idioms.

Idiom completion as an NLP task

The idioms are split into idiom starts (the first N-1 words) and accepted completions (a list of possible last words to complete the idiom). Of the 3259 rows, there are 154 where there are more than one accepted completion.

This dataset can be used to measure a generative language models' ability to complete well known idioms, or as a masked language modelling task.

Using the dataset

Loading the dataset with original idioms only:

from datasets import load_dataset

ds = load_dataset("Sprakbanken/Norwegian_idioms", split="test")

Loading the dataset with translated idioms included:

from datasets import load_dataset

ds = load_dataset("Sprakbanken/Norwegian_idioms", split="test", name="include_translated_idioms")

Idiom frequencies

idiom_freqs/idiom_frequencies.csv contain the idiom, language code and frequency in the online library for all the original idioms
idiom_freqs/translated_idiom_frequencies.csv contain the idiom, language code, frequency in the online library, source idiom and source language code for all the translated idioms

Idiom graphs

There is considerable linguistic overlap between the idioms. For example, though there are 3553 unique idioms, there are only 803 unique starter words.

We have arranged the idioms as trees, where the roots are the start words, and the idioms are found by following a path from a root node to a leaf node.
Example:

"alt": {
    "er": {
        "bare": {
            "fryd": {
                "og": {
                    "gammen": {}
                }
            }
        },
        "klappet": {
            "og": {
                "klart": {}
            }
        },
        "såre": {
            "vel": {}
        }
    },
    "går": {
        "sin": {
            "vante": {
                "gang": {}
            }
        }
    },
}

Here is a tree where "alt" is the root, the idioms "alt er bare fryd og gammen", "alt er klappet og klart", "alt er såre vel" and "alt går sin vante gang" can be found by traversing the tree.
These trees can be found in idiom_graphs/.

There are also flat versions of the trees (where consecutive nodes/words with only one child are merged into word sequences)

Example:

"alt": {
    "er": {
        "bare fryd og gammen": "",
        "klappet og klart": "",
        "såre vel": ""
    },
    "går sin vante gang": "",
}