Norwegian Petroleum Corpus

Research
Data Quality
Data-Centric ML
Machine Translation
RAG
Norwegian Petroleum Corpus

52,865

MT pairs

73,866

RAG chunks

0

Split leakage

Tech Stack

Python
Web Crawling
PDF Processing
LaBSE
Multilingual NLP
Information Retrieval

Description

This project develops an open-source and reproducible pipeline for constructing Norwegian petroleum-domain research data from public English and Norwegian documents. The current implementation processes Equinor web pages and PDFs to support machine translation, language-model research, and information retrieval while preserving document-level provenance.

    My Contributions

    As the project lead, I designed and implemented the complete corpus-construction pipeline, from source collection and text extraction to bilingual document pairing, sentence alignment, quality filtering, and dataset splitting. I also created traceable research outputs for machine translation, language-model research, and information retrieval.

    Project Highlights

    Reproducible Corpus Pipeline

    Norwegian Petroleum Corpus
    1. 1

      Collect and Extract

      The pipeline collects public English and Norwegian web pages and PDFs from Equinor. It extracts the text and saves the title, language, URL, document ID, and other source information.

    2. 2

      Pair Bilingual Documents

      For the machine translation data, official language-switch links are used to identify matching English and Norwegian articles. Reciprocal links are checked to confirm that the two articles form a valid bilingual pair.

    3. 3

      Align Sentences

      The matched articles are divided into sentences. LaBSE is then used to find English and Norwegian sentences that are likely to be translations of each other.

    4. 4

      Filter Sentence Pairs

      The pipeline retains sentence pairs with similar meanings and reasonable lengths. Website noise and duplicate pairs are removed, and the Norwegian sentences are checked to ensure that they are written in Bokmal.

    5. 5

      Create Dataset Splits

      The final sentence pairs are divided into training, validation, and test sets by bilingual document pair. This prevents content from the same article from appearing in different sets.

    PDFs and web pages without a matching translation are used only for the language-model and retrieval data.

    Research Outputs

    OutputIntended use
    English-Norwegian sentence pairsMachine translation
    Monolingual document recordsLanguage-model training and adaptation
    Traceable text chunksInformation retrieval and RAG

    Research Connection

    This corpus provides the new-domain data for a subsequent continual-adaptation study of new-source learning, old-source forgetting, and replay-based retention in petroleum-domain machine translation.

    View Continual Adaptation Study