abhinandansamal/bidirectional_odia_german_translation_parallel_corpus

Dataset Card for Odia-German Parallel Corpus (Odia-Deu-News)

---

Dataset Summary

The Odia-German Parallel News Corpus is a bidirectional dataset designed to enhance contextual understanding in low-resource language translation. It was developed as part of the Master's Thesis "Enhancing Contextual Understanding in Low-Resource Languages Using Multilingual Transformers" at IU International University of Applied Sciences, Germany.

The corpus consists of 7,352 training instances (derived from 3,676 unique parallel pairs) sourced from contemporary Odia news articles published between January 2024 and June 2025. Unlike standard sentence-aligned corpora, this dataset preserves local context by frequently grouping 2-3 related sentences per line, making it particularly suitable for document-level or context-aware translation tasks.

Supported Tasks and Leaderboards

• Machine Translation (MT): The dataset is formatted for Odia -> German and German -> Odia translation.
• Multitask Learning: The data is pre-processed with task-specific prefixes to train a single model on both translation directions simultaneously.

Languages

• Source/Target: Odia (Oriya) - ory_Orya (Indo-Aryan language spoken in Odisha, India).
• Source/Target: German - deu_Latn (West Germanic language).

---

Dataset Structure

Data Instances

The dataset is provided in a JSON Lines (.jsonl) format. To facilitate bidirectional training, every unique parallel pair appears twice: once with Odia as the input and once with German as the input.

Example Entry (Odia -> German):

{
  "id": 1801,
  "original_id": 901,
  "URL": "https://www.dharitri.com/forest-thought/",
  "domain": "Editorial",
  "topic": "ଜଙ୍ଗଲ ଚିନ୍ତା",
  "publication_date": "May, 2025",
  "input_text": "translate Odia to German: ଏହି ପରିପ୍ରେକ୍ଷୀରେ ସରକାର ଓ କମ୍ପାନୀ ସହିତ ପ୍ରତ୍ୟେକ ବ୍ୟକ୍ତିଙ୍କୁ ସଚେତନ ହେବାକୁ ପଡ଼ିବ। ନିଜେ ନିଜ ସ୍ତରରେ ଅଧିକ ଚାରା ରୋପଣ କରିପାରିଲେ ନିଜ ପରିବେଷ୍ଟନୀ ସୁରକ୍ଷିତ ରହିପାରିବ।",
  "target_text": "In diesem Zusammenhang muss sich jeder Einzelne, einschließlich der Regierung und der Unternehmen, der Situation bewusst sein. Durch das Pflanzen von mehr Bäumen auf ihrem eigenen Boden kann die Umwelt geschützt werden."
}

Example Entry (German -> Odia):

{
  "id": 854,
  "original_id": 427,
  "URL": "https://www.dharitri.com/economy-of-these-15-countries-is-growing-faster-than-india-the-country-on-top-is-list-includes/",
  "domain": "International, National, Trade",
  "topic": "ଏହି ୧୫ଟି ଦେଶର ଅର୍ଥନୀତି ଭାରତଠାରୁ ତୀବ୍ର ଗତିରେ ବୃଦ୍ଧି ପାଉଛି, ଶୀର୍ଷରେ ଥିବା ଦେଶ ହେଉଛି…",
  "publication_date": "April, 2025",
  "input_text": "translate German to Odia: Senegals Wirtschaft wird in diesem Jahr voraussichtlich um 9,3 Prozent wachsen.",
  "target_text": "ସେନେଗାଲର ଅର୍ଥନୀତି ଚଳିତ ବର୍ଷ ୯.୩ ପ୍ରତିଶତ ହାରରେ ବୃଦ୍ଧି ପାଇବ ବୋଲି ଆଶା କରାଯାଉଛି।"
}

Data Fields

• id: Unique identifier for the training instance (1 to 7352).
• original_id: Identifier linking back to the original unique parallel pair (1 to 3676).
• URL: The source URL of the news article (placeholder or actual).
• domain: The source newspaper domain (e.g., Dharitri, Sambad).
• topic: The news category (e.g., national-news, sports, science).
• publication_date: Date of article publication.
• input_text: The string to be fed into the model. Includes the task prefix:
  • "translate Odia to German: "
  • "translate German to Odia: "
• target_text: The expected translation output.

Dataset Creation

Curation Rationale

Low-resource languages like Odia often lack high-quality, contemporary parallel data for German. Existing datasets often rely on religious texts or outdated crawls. This dataset bridges that gap by providing high-quality news data with manually verified translations.

Source Data

The text was sourced from the online editions of two prominent Odia daily newspapers:

• Dharitri: https://www.dharitri.com/
• Sambad: https://sambad.in/

Data Processing & Formatting

1. Web Scraping: Articles were scraped using a custom Python pipeline (requests, newspaper3k, BeautifulSoup).

2. Odia Filtering: A custom filter filter_odia_text was applied to retain only characters in the Unicode range U+0B00 – U+0B7F (Odia script) and specific punctuation, removing all English boilerplate and navigation elements.

3. Structuring: The parallel text files were merged into a structured JSONL format, ensuring line-by-line alignment.

4. Bidirectional Transformation: The dataset was expanded by duplicating instances and adding directional prefixes (translate Odia to German: / translate German to Odia:), then shuffled to prevent ordering bias during training.

Annotations & Quality Control

A hybrid quality control process was implemented to ensure accuracy:

• Odia Source Validation: 100% of the 3,676 Odia lines were manually reviewed and corrected by a native Odia speaker.
• German Target Validation:
  • Gold Standard (2,000 lines): Manually evaluated and corrected by a native German speaker.
  • Silver Standard (1,676 lines): Generated via machine translation and subsequently post-edited by the author. A primary focus was the manual localization of Indian numbering terms (e.g., "Lakh", "Crore") into German equivalents (Millions, Billions).

Considerations for Using the Data

Social Impact of Dataset

This dataset contributes to the preservation and technological inclusion of the Odia language (spoken by over 40 million people). It facilitates cross-cultural communication and information access between Odisha and German-speaking regions.

Discussion of Biases

• Domain Bias: The vocabulary is heavily skewed towards formal, journalistic reporting (politics, sports, trade) and may not reflect colloquial speech.
• MT Artifacts: The "Silver Standard" portion of the German data was originally generated by MT. Despite human post-editing, some structural artifacts may remain.

Additional Information

Dataset Curators

Abhinandan Samal (Master's Student, IU International University of Applied Sciences).

Licensing Information

This dataset is released under the Creative Commons Attribution 4.0 International (CC BY 4.0) license.

Note: While the parallel pairings and annotations are the intellectual property of the curator, the underlying news content remains the copyright of the respective publications (Dharitri and Sambad). This dataset is intended for academic and research purposes.

Citation Information

If you use this dataset in your research, please cite the associated Thesis:

@mastersthesis{samal2025enhancing,
  title={Enhancing Contextual Understanding in Low-Resource Languages Using Multilingual Transformers},
  author={Samal, Abhinandan},
  school={IU International University of Applied Sciences},
  year={2025},
  address={Germany},
  note={Dataset: Odia-German Parallel News Corpus}
}