README.md

---
task_categories:
- translation
language:
- pl
- cs
- de
---
</h1>

<div class="section">
<h2>1. Dataset Summary</h2>

<p>
This dataset provides parallel corpora for pivot-based neural machine translation research.
It contains Polish–Czech, Czech–German, and Polish–German parallel sentence pairs,
derived from the JRC-Acquis corpus and preprocessed for reproducible experiments.
</p>

<p>
The dataset was created to support research on pivot translation,
particularly investigating Czech as an intermediate language for Polish–German translation.
</p>
</div>

<div class="section">
<h2>2. Languages</h2>

<ul>
<li>Polish (pl)</li>
<li>Czech (cs)</li>
<li>German (de)</li>
</ul>

<p><strong>Language Pairs:</strong></p>
<ul>
<li>pl–cs</li>
<li>cs–de</li>
<li>pl–de</li>
</ul>
</div>

<div class="section">
<h2>3. Dataset Structure</h2>

<p>
Each language pair is split into training, validation, and test sets
with the following ratio:
</p>

<pre>9 : 0.5 : 0.5</pre>

<p>
Directory structure:
</p>

<pre>
./
  ├ JRC-Acquis.cs-de.*
  ├ JRC-Acquis.cs-pl.*
  └ JRC-Acquis.de-pl.*
</pre>

<p>
Each file contains aligned parallel sentence pairs.
</p>
</div>

<div class="section">
<h2>4. Data Source</h2>

<p>
The dataset is derived from the <strong>JRC-Acquis Parallel Corpus</strong>,
which consists of European Union legislative texts.
</p>

<p>
The original corpus is widely used in multilingual NLP research.
</p>
</div>

<div class="section">
<h2>5. Preprocessing</h2>

<p>The following preprocessing steps were applied:</p>

<ul>
<li>Sentence alignment verification</li>
<li>Basic normalization</li>
<li>Train/validation/test splitting (9:0.5:0.5)</li>
</ul>

<p>
No subword segmentation (e.g., BPE or SentencePiece) is applied in this dataset version.
</p>
</div>

<div class="section">
<h2>6. Intended Use</h2>

<p>This dataset is intended for:</p>

<ul>
<li>Pivot-based neural machine translation research</li>
<li>Low-resource translation experiments</li>
<li>Comparative evaluation of pivot language selection</li>
<li>Evaluation using BLEU, GLEU, and CHRF metrics</li>
</ul>

<p>
It is not specifically designed for commercial deployment.
</p>
</div>

<div class="section">
<h2>7. Reproducibility</h2>

<p>
The dataset includes pre-split files identical to those used in the original experiments.
</p>

<p>
In the original study:
</p>

<ul>
<li>60,000 sentence pairs were used for training per translation model</li>
<li>10,000 sentences were used for testing</li>
<li>Transformer-based models were applied</li>
</ul>
</div>

<div class="section">
<h2>8. Evaluation Metrics</h2>

<ul>
<li>BLEU</li>
<li>GLEU</li>
<li>CHRF</li>
</ul>
</div>

<div class="section">
<h2>9. Ethical Considerations</h2>

<p>
The dataset is based on publicly available legislative texts from the European Union.
No personally identifiable information (PII) is intentionally included.
</p>

<p>
Users should verify compliance with the original JRC-Acquis license before use.
</p>
</div>

<div class="section">
<h2>10. Citation</h2>

<p>If you use this dataset, please cite:</p>

<pre>
Tokunaga, N., Ninomiya, T., & Tamura, A. (2020).
Neural Machine Translation for Polish–German via Czech as a Pivot Language.
</pre>

<p>Dataset citation:</p>

<pre>
Tokunaga, N., Ninomiya, T., & Tamura, A. (2020).
TransformerPivotTranslation Dataset.
Hugging Face.
https://huggingface.co/datasets/SHSK0118/TransformerPivotTranslation
</pre>

</div>

<div class="section">
<h2>11. License</h2>

<p>
This dataset inherits the licensing conditions of the original JRC-Acquis corpus.
Users must ensure compliance with the original terms of use.
</p>
</div>

<div class="section">
<h2>12. Contact</h2>

<p>
For academic inquiries, please contact the dataset authors.
</p>
</div>