Hugging Face's logo

Datasets:
abhinandansamal
/
bidirectional_odia_german_translation_parallel_corpus

Tasks:
Translation
Languages:
Oriya
German
Size:
1K<n<10K
Tags:
translation
machine-translation
parallel-corpus
news
journalism
odia
License:
Dataset card Data Studio Files Community

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Odia-German Parallel Translation Corpus

Dataset Description

This repository hosts a meticulously curated parallel corpus for Machine Translation (MT) research between Odia (or) and German (de). It comprises two primary files:

  • authentic_corpus_final.jsonl: The main, highly curated parallel corpus, containing original source information and parallel sentences.
  • bidirectional_corpus_final.jsonl: A derivative of the authentic corpus, specifically formatted for bidirectional fine-tuning of neural machine translation models, with task-specific prefixes.

The dataset is designed to support the development of NMT systems, particularly for under-resourced language pairs like Odia-German, and to facilitate research into data curation strategies under resource constraints.

Dataset Structure

The data files are in JSON Lines (.jsonl) format, where each line is a JSON object. Both files contain the same number of fields, but with different naming conventions and content for the primary text fields, reflecting their distinct purposes.

authentic_corpus_final.jsonl

This file serves as the main, raw parallel corpus. Each object represents one parallel line pair.

Example Structure:

{
  "id": 1,
  "URL": "[https://www.dharitri.com/india-needs-8-9-growth-to-become-viksit-bharat-by-2047-raghuram-rajan/](https://www.dharitri.com/india-needs-8-9-growth-to-become-viksit-bharat-by-2047-raghuram-rajan/)",
  "domain": "National, Trade",
  "topic": "୨୦୪୭ ସୁଦ୍ଧା କେମିତି ବିକଶିତ ହେବ ଭାରତ? ବଡ ବୟାନ ଦେଲେ ପୂର୍ବତନ ଆରବିଆଇ ଗଭର୍ଣ୍ଣର",
  "publication_date": "June,2025",
  "sentence_ory_Orya": "ପୂର୍ବତନ ଆରବିଆଇ ଗଭର୍ଣ୍ଣର ତଥା ପ୍ରଖ୍ୟାତ ଅର୍ଥନୀତିଜ୍ଞ ରଘୁରାମ ରାଜନ ଏକ ସାକ୍ଷାତକାରରେ କହିଛନ୍ତି, “ମୁଁ ବହୁ ଦିନ ଧରି ଜୋର ଦେଇ କହୁଛି—ଯଦି ଆମେ ୨୦୪୭ ସୁଦ୍ଧା ବିକଶିତ ଦେଶ ହେବାକୁ ଚାହୁଁ, ତେବେ ଆମକୁ ୮, ୮.୫ ରୁ ୯% ଅଭିବୃଦ୍ଧି ଆବଶ୍ୟକ, କାରଣ ଆମେ ଏବେ ବି ଏକ ତୁଳନାତ୍ମକ ଭାବେ ଦରିଦ୍ର ରାଷ୍ଟ୍ର।”",
  "sentence_deu_Latn": "Der ehemalige Gouverneur der Reserve Bank of India (RBI) und renommierte Ökonom Raghuram Rajan sagte in einem Interview: „Ich betone schon seit langem, dass wir ein Wachstum von 8, 8,5 bis 9 Prozent brauchen, wenn wir bis 2047 ein entwickeltes Land werden wollen, denn wir sind immer noch ein vergleichsweise armes Land.“"
}

Fields:

  • id (int): A unique identifier for the line pair within this file.
  • URL (string): The URL of the news article from which the line was extracted.
  • domain (string): Categories indicating the topic domain(s) of the news content (e.g., "National", "Trade").
  • topic (string): A placeholder for Odia text, which likely contains a short description or headline related to the topic in Odia.
  • publication_date (string): The publication month and year of the original news article.
  • sentence_ory_Orya (string): The text in Odia (Oriya script).
  • sentence_deu_Latn (string): The corresponding text in German (Latin script).

bidirectional_corpus_final.jsonl

This file contains transformed versions of the authentic_corpus_final.jsonl entries, typically duplicated and prefixed for bidirectional training of MT models.

{
  "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/](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।"
}

Fields:

  • id (int): A unique identifier for the training instance within this file (different from original_id).
  • original_id (int): Links back to the id field in the authentic_corpus_final.jsonl file, indicating the original parallel pair.
  • URL (string): The URL of the news article from which the line was extracted.
  • domain (string): Categories indicating the topic domain(s) of the news content (e.g., "National", "Trade").
  • topic (string): A placeholder for Odia text, which likely contains a short description or headline related to the topic in Odia.
  • publication_date (string): The publication month and year of the original news article.
  • input_text (string): The source text for the translation task, typically including a task-specific prefix (e.g., "translate German to Odia:").
  • target_text (string): The target text for the translation task.

Data Collection and Curation

The dataset began with the collection of 3,676 unique parallel line pairs extracted from a total of 298 news articles sourced from two prominent Odia newspapers: "Dharitri" and "Sambad". Each "line" was designed to provide contextual information, often containing 2-3 sentences, although single-sentence lines are also present. The corpus encompasses a diverse range of news domains, including:

  • National
  • International
  • Lifestyle
  • Sports
  • Trade
  • Environmental
  • Science and Technology
  • Leisure
  • Commerce
  • Metro
  • State
  • Editorial

A multi-faceted quality control and curation process was applied:

  • Odia Corpus Validation: All 3,676 lines on the Odia side of the parallel corpus underwent thorough evaluation and validation by a native Odia speaker (the dataset creator (me)), ensuring high linguistic fidelity.

  • German Corpus Curation:

    • A high-quality subset of 2,000 German lines (corresponding to 2,000 of the original parallel pairs) was meticulously human-evaluated and corrected by a native German speaker, forming a core validated segment.
    • The remaining 1,676 German lines were initially generated using Google Translate. These machine-translated lines were incorporated to broaden the corpus's vocabulary and grammatical structures, supplementing the human-validated content.

The bidirectional_corpus_final.jsonl was derived from the authentic_corpus_final.jsonl by creating two training instances for each original pair, effectively doubling the lines to 7,352 distinct instances for bidirectional training.

The size of this dataset was primarily dictated by the memory and computational constraints of the available training environment (NVIDIA T4 on Google Colab Pro), necessitating a focus on quality and strategic augmentation within these limits.

Dataset Curators

  • Odia Text Validation: Abhinandan Samal (dataset creator/author)
  • German Text Validation: Sarah Papperitz

Usage

You can load this dataset directly using the Hugging Face datasets library.

from datasets import load_dataset

# To load the full authentic corpus
authentic_corpus = load_dataset("abhinandansamal/bidirectional_odia_german_translation_parallel_corpus", data_files="authentic_corpus_final.jsonl")
print(authentic_corpus)

# To load the bidirectional fine-tuning corpus
bidirectional_corpus = load_dataset("abhinandansamal/bidirectional_odia_german_translation_parallel_corpus", data_files="bidirectional_corpus_final.jsonl")
print(bidirectional_corpus)

# Example: Accessing an entry from the authentic corpus
# print(authentic_corpus['train'][0]) # 'train' is the default split name if not specified

Citation

If you use this dataset in your research, please consider citing it:

@misc{SamalThesis2025,
  author = Abhinandan Samal,
  title  = Odia-German Parallel Translation Corpus,
  school = IU International University of Applied Sciences,
  publisher = Hugging Face,
  year   = 2025,
  url = {[https://huggingface.co/datasets/abhinandansamal/bidirectional_odia_german_translation_parallel_corpus](https://huggingface.co/datasets/abhinandansamal/bidirectional_odia_german_translation_parallel_corpus)}
}

License

This dataset is released under the Apache license 2.0 License.

Limitations and Bias

  • Domain Specificity: While covering various news sub-domains, the dataset is primarily composed of journalistic text. Its utility and performance for tasks in vastly different domains (e.g., legal, medical, literary, informal conversations) may be limited.
  • Source Bias: The content inherits topical and stylistic biases inherent to its sources (Dharitri and Sambad newspapers).
  • Machine Translation Influence: The inclusion of Google Translate-generated German lines may introduce or reinforce specific stylistic patterns or errors characteristic of machine translation systems. Users should be aware of this potential bias, especially for applications requiring very high-fidelity human-like translation.
  • Size: As a relatively small corpus, it may not cover the full breadth of linguistic phenomena required for highly robust, general-purpose MT systems.
Downloads last month
0