dsfsi/govza-sa-cabinet-statements-sentence-aligned

Gov-ZA Multilingual Cabinet Statements (Sentence-Aligned)

Dataset Description

This dataset contains sentence-aligned parallel text from South African government cabinet statements in 11 official languages. The data is sourced from the Government Communication and Information System (GCIS) and scraped from www.gov.za/cabinet-statements.

Key Features:

• 📊 55 language pair combinations covering 11 South African languages
• 🔗 Sentence-level alignment using LASER embeddings
• 📈 Alignment confidence scores (cosine similarity)
• 🎯 Train/Test/Eval splits for each language pair
• 🌍 Low-resource African languages represented

Supported Languages

Language	Code	Language	Code
Afrikaans	afr	Sepedi (Northern Sotho)	nso
English	eng	Setswana	tsn
isiNdebele	nbl	Siswati	ssw
isiXhosa	xho	Tshivenda	ven
isiZulu	zul	Xitstonga	tso
Sesotho	sot

Dataset Structure

Data Instances

Each instance contains a pair of aligned sentences with an alignment confidence score:

{
  "afr": "Die Kabinet het sy kommer uitgespreek oor die onlangse arbeidsonrus...",
  "eng": "Cabinet expressed concern at recent labour unrest at Lonmin's Marikana Mine...",
  "score": 0.8649104,
  "__index_level_0__": 2
}

Data Fields

• {lang1} (string): Source language sentence
• {lang2} (string): Target language sentence (aligned translation)
• score (float): Alignment confidence score (0.0 to 1.0) based on cosine similarity of LASER embeddings
• __index_level_0__ (int): Original index in the full dataset

Data Splits

Each language pair configuration has three splits:

Split	Purpose	Size
train	Training	~70%
test	Testing	~15%
eval	Evaluation	~15%

Dataset Creation

Source Data

The dataset is created from South African government cabinet statements published on the official government website. These statements are professionally translated into all 11 official languages, making them a high-quality source for parallel text.

Sentence Alignment Process

1. Scraping: Cabinet statements scraped from gov.za
2. Tokenization: Sentence tokenization using NLTK
3. Preprocessing: Language-specific text cleaning (URLs, titles, special characters)
4. Embedding: Sentence embeddings generated using LASER (Language-Agnostic SEntence Representations)
5. Alignment: Sentences aligned using cosine similarity of embeddings
6. Filtering: Alignment pairs filtered by confidence score threshold (>= 0.65)

Usage

Loading the Dataset

from datasets import load_dataset

# Load specific language pair
dataset = load_dataset("dsfsi/govza-sa-cabinet-statements-sentence-aligned", "afr-eng")

# Access splits
train_data = dataset["train"]
test_data = dataset["test"]
eval_data = dataset["eval"]

# Iterate through examples
for example in train_data:
    source = example["afr"]
    target = example["eng"]
    confidence = example["score"]
    print(f"Score: {confidence:.2f}")
    print(f"AFR: {source}")
    print(f"ENG: {target}\n")

Available Language Pair Configurations

The dataset includes the following configurations (language pairs):

Click to see all 55 language pair combinations

Afrikaans pairs: afr-eng, afr-nbl, afr-nso, afr-sot, afr-ssw, afr-tsn, afr-tso, afr-ven, afr-xho, afr-zul

English pairs: eng-nbl, eng-nso, eng-sot, eng-ssw, eng-tsn, eng-tso, eng-ven, eng-xho, eng-zul

isiNdebele pairs: nbl-nso, nbl-sot, nbl-ssw, nbl-tsn, nbl-tso, nbl-ven, nbl-xho, nbl-zul

Sepedi pairs: nso-sot, nso-ssw, nso-tsn, nso-tso, nso-ven, nso-xho, nso-zul

Sesotho pairs: sot-ssw, sot-tsn, sot-tso, sot-ven, sot-xho, sot-zul

Siswati pairs: ssw-tsn, ssw-tso, ssw-ven, ssw-xho, ssw-zul

Setswana pairs: tsn-tso, tsn-ven, tsn-xho, tsn-zul

Xitstonga pairs: tso-ven, tso-xho, tso-zul

Tshivenda pairs: ven-xho, ven-zul

isiXhosa-isiZulu: xho-zul

Example: Training a Translation Model

from datasets import load_dataset
from transformers import MarianMTModel, MarianTokenizer

# Load data
dataset = load_dataset("dsfsi/govza-sa-cabinet-statements-sentence-aligned", "afr-eng", split="train")

# Your training code here
# ...

Alignment Quality

Alignment pairs are filtered by cosine similarity score >= 0.65. The distribution of alignment scores varies by language pair:

• High-resource pairs (e.g., eng-afr): Typically higher alignment scores (0.75-0.95)
• Low-resource pairs (e.g., ven-nbl): More varied scores (0.65-0.85)

Higher scores indicate stronger semantic similarity between aligned sentences.

Important: Data Quality Recommendations

⚠️ This dataset provides valuable parallel data for language pairs that are otherwise difficult to obtain, but it is not bulletproof. Please read these recommendations carefully before using the data.

Recommended Quality Assurance Steps

Before using this dataset for your specific use case, we strongly recommend:

1. Evaluate optimal filtering thresholds for your language pair

   • The current filtering threshold (cosine similarity >= 0.65) is a general baseline
   • Different language pairs may benefit from different thresholds depending on your quality requirements
   • We recommend analyzing the score distribution for your chosen language pair(s) and determining an appropriate threshold for your application
   • Consider filtering by higher scores (e.g., >= 0.75 or >= 0.80) if you need higher precision, especially for low-resource language pairs

2. Verify alignments with alternative methods

   • Cross-check alignment quality using other sentence alignment tools (e.g., AWESOME-align, SimAlign, Vecalign)
   • Consider using different cross-lingual embeddings (e.g., LaBSE, mBERT, XLM-RoBERTa) to validate alignment quality
   • Manual inspection of a sample of alignments for your target language pair is highly recommended

3. Understand the limitations

   • Alignment quality varies significantly across language pairs
   • Some language pairs may have systematic alignment errors due to differences in sentence structure or translation style
   • Low-resource language pairs may have fewer training examples and more noise

Why This Matters

This dataset was created to provide parallel text data for South African language pairs that are otherwise extremely difficult to obtain. While the alignment process uses state-of-the-art methods (LASER embeddings), no automated alignment system is perfect. The quality you need depends on your specific use case:

• For exploratory research or initial experiments: The default threshold (>= 0.65) may be sufficient
• For training production translation models: Consider higher thresholds (>= 0.75) and additional validation
• For linguistic analysis or evaluation benchmarks: Manual verification of alignments is strongly recommended

The value of this dataset lies in making available parallel sentences for language pairs where such data is scarce, but users should apply appropriate quality controls for their specific needs.

Considerations for Using the Data

Social Impact of Dataset

This dataset supports:

• ✅ Development of translation systems for South African languages
• ✅ Research on low-resource language NLP
• ✅ Preservation and promotion of linguistic diversity
• ✅ Improved access to government information across language communities

Limitations

• Domain-specific: Government/political domain - may not generalize to other domains
• Formal register: Professional translations in formal style
• Alignment quality varies: Some language pairs have fewer high-quality alignments
• Historical context: Statements reflect specific time periods and political contexts

Additional Information

Dataset Curators

• Organization: Data Science for Social Impact (DSFSI) Research Group
• Institution: University of Pretoria
• Lead: Vukosi Marivate (@vukosi)

Licensing Information

• Data License: CC BY 4.0
• Code License: MIT License

Citation Information

Paper:

@inproceedings{lastrucci-etal-2023-preparing,
    title = "Preparing the Vuk{'}uzenzele and {ZA}-gov-multilingual {S}outh {A}frican multilingual corpora",
    author = "Richard Lastrucci and Isheanesu Dzingirai and Jenalea Rajab and Andani Madodonga and Matimba Shingange and Daniel Njini and Vukosi Marivate",
    booktitle = "Proceedings of the Fourth workshop on Resources for African Indigenous Languages (RAIL 2023)",
    month = may,
    year = "2023",
    address = "Dubrovnik, Croatia",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2023.rail-1.3",
    pages = "18--25"
}

Dataset:

@dataset{marivate_vukosi_2023_7635168,
  author       = {Marivate, Vukosi and
                  Shingange, Matimba and
                  Lastrucci, Richard and
                  Dzingirai, Isheanesu and
                  Rajab, Jenalea},
  title        = {The South African Gov-ZA multilingual corpus},
  month        = feb,
  year         = 2023,
  publisher    = {Zenodo},
  version      = {1.0},
  doi          = {10.5281/zenodo.7635168},
  url          = {https://doi.org/10.5281/zenodo.7635168}
}

Disclaimer

This dataset contains machine-readable data extracted from online cabinet statements from the South African government, provided by the Government Communication Information System (GCIS). While efforts were made to ensure the accuracy and completeness of this data, there may be errors or discrepancies between the original publications and this dataset. No warranties, guarantees or representations are given in relation to the information contained in the dataset.

Contributions

Thanks to the DSFSI team for creating this dataset.

Links

• 🏠 Homepage: https://github.com/dsfsi/gov-za-multilingual
• 📄 Paper: https://arxiv.org/abs/2303.03750
• 🗂️ Source Code: https://github.com/dsfsi/gov-za-multilingual
• 📊 Zenodo: https://doi.org/10.5281/zenodo.7635168