Datasets:

UppsalaNLP
/

sou-corpus

license: cc-by-4.0
language:
  - sv
task_categories:
  - text-generation
  - token-classification
tags:
  - swedish
  - government-reports
  - dependency-parsing
  - universal-dependencies
  - nlp
size_categories:
  - 100K<n<1M
source_datasets:
  - original

SOU Corpus - Swedish Government Official Reports

Cleaned and dependency-parsed Swedish Government Official Reports (Statens offentliga utredningar) from 1994-2020.

Dataset Description

This dataset contains sentence-segmented and dependency-parsed text from Swedish Government Official Reports. The original documents were cleaned, processed, and annotated with Universal Dependencies-style parsing.

Fields

document_id: Original document identifier (can be linked to Riksdagen open data)
text_type: Type of text section
- full_text: Main report body
- summary_swedish: Standard Swedish summary
- summary_simple_swedish: Simple Swedish (lättläst) summary
- summary_english: English summary
section: Section header from the document
text: Plain text sentence
parsed: Dependency-parsed sentence (token//POS//deprel//head format)

Parsed Format

Each token in the parsed field follows the format:

word//POS_TAG//DEPENDENCY_RELATION//HEAD_INDEX

Example:

Sverige//PM|NOM//nsubj//3 är//VB|PRS|AKT//cop//3 ett//DT|NEU|SIN|IND//det//3 land//NN|NEU|SIN|IND|NOM//ROOT//3

Usage

from datasets import load_dataset

dataset = load_dataset("UppsalaNLP/sou-corpus")

# Access train/test splits
train = dataset["train"]
test = dataset["test"]

# Example
print(train[0]["text"])
print(train[0]["section"])
print(train[0]["text_type"])

Extract Tokens

def parse_tokens(parsed_str):
    tokens = []
    for t in parsed_str.split(' '):
        parts = t.split('//')
        if len(parts) >= 4:
            tokens.append({
                'word': parts[0],
                'pos': parts[1],
                'deprel': parts[2],
                'head': int(parts[3]) if parts[3].isdigit() else 0
            })
    return tokens

tokens = parse_tokens(train[0]["parsed"])

Source

Documents obtained from Riksdagens öppna data. Original document URLs follow the pattern: https://data.riksdagen.se/dokument/{document_id}.html

Citation

@inproceedings{durlich-etal-2022-cause,
    title = "Cause and Effect in Governmental Reports: Two Data Sets for Causality Detection in Swedish",
    author = "D{\"u}rlich, Luise and Reimann, Sebastian and Finnveden, Gustav and Nivre, Joakim and Stymne, Sara",
    booktitle = "Proceedings of the First Workshop on Natural Language Processing for Political Sciences",
    month = jun,
    year = "2022",
    address = "Marseilles, France"
}

License

This dataset is licensed under CC BY 4.0.