Korean PDF Annotation Specialist

Mercor; Helsinki; On-site; EUR 32,000 - 52,000; Full time

Job summary

Obsidian is seeking a meticulous Korean-language annotator to create accurate transcription and structural mappings for diverse Korean documents. You will bound regions, assign types, and ensure reading order aligns with multi-column layouts, including handwritten content.

The role emphasizes character-level accuracy, reviewer-based quality checks, and handling complex layouts with real-world documents. Strong Hangul competency is essential for this work.

Qualifications

• Native Korean speaker with Hangul proficiency and accuracy essential.
• Experience in bilingual transcription, editorial work, or AI training data is preferred.
• Exact, character-level transcription and strict consistency across pages required.

Responsibilities

• Source Korean PDFs and identify multimodal elements.
• Annotate page structure: title, headings, paragraphs, lists, tables, figures, captions, formulas, questions, answers.
• Record relationships between regions and parent figures/tables.
• Transcribe faithfully in Hangul, including hanja, marking unreadable regions.
• Capture metadata: language, document type, source, page dimensions, and handwriting flags.
• Review work through second Korean expert feedback and ensure layout diversity.

Skills

Native Korean speaker

Strong attention to detail

Bilingual transcription/editing

Editorial/rating experience

Job description

Fluent Language Skills Required: Korean. Native fluency in Korean, including full command of Hangul, is required for this position. All annotation and transcription work is performed in Korean.

Why This Role Exists

Document understanding breaks down fastest in the languages that parsing and vision-language models rarely see. This project builds training data for exactly those languages: Korean, alongside Japanese and five Indic scripts. Each task takes a real, publicly available PDF page and produces a complete structural map of that page, paired with a faithful transcription of every text region in the original script.

The dataset deliberately concentrates on the material models handle worst: handwriting, dense multi-column layouts, tables, diagrams, and mixed-script pages. Documents are drawn from newspapers, textbooks, examinations, and everyday formats such as flyers, forms, manuals, menus, brochures, notices and worksheets, so that the corpus reflects the real diversity of Korean documents rather than a narrow band of easily parsed ones.

Delivered work is human-authored throughout. Component identification, component typing, reading order and all transcription are performed by people, not generated by parsing models.

What You'll Do

• Source documents: find a publicly accessible Korean PDF in an assigned document type, containing at least one multimodal element (images, tables, diagrams, or handwriting), and record where you obtained it

• Annotate structure: identify and bound every meaningful region of the page - document title, section heading, paragraph, list, table, figure, diagram, caption, formula, question, answer field - and assign each a component type and a reading-order index

• Record relationships: link each region to the figure or table it belongs to through a parent component identifier

• Transcribe faithfully: reproduce all text exactly as it appears in Hangul, including any hanja and handwritten content, flagging any region where the source is not legible

• Capture page metadata: language, document type, source, page dimensions, and flags for tables, formulas and handwriting

• Review a colleague's work: every task is reviewed end to end by a second Korean expert, and experienced annotators take on that review

Who You Are

• You are a native Korean speaker with full command of Hangul, including hanja where it appears in older or formal documents

• You have worked in bilingual transcription, translation, editorial work, or AI training data, ideally with reviewer experience

• You are exact: character-level accuracy matters more here than speed, and a single wrong jamo is a defect

• You are systematic: you apply a taxonomy consistently across hundreds of pages rather than improvising per document

• You are comfortable with unfamiliar layouts: multi-column newspapers, exam papers, handwritten forms

Nice-to-Have Specialties

• AI training data: annotation, labeling, grading, or bilingual evaluation for training datasets

• Transcription and localization: MTPE, subtitling, bilingual QA, OCR correction or post-editing

• Document production: typesetting, copy-editing, proofreading, or digitization of Korean-language material

• Script and encoding: Unicode normalization, Korean input methods, Hangul jamo composition, and hanja handling

What Success Looks Like

• Every meaningful region on the page is captured, correctly bounded and correctly typed

• Reading order reflects how the page is actually read, including across columns

• Transcriptions match the source character for character, in Hangul rather than romanization

• Your tasks pass second-expert review the first time

• The documents you bring in add layout diversity rather than repeating templates already in the corpus

Why Join Mercor

• Build the training data that makes document AI work in scripts it currently handles badly

• Work from real published Korean documents rather than synthetic or templated pages

• Quality leads on this project: accuracy is the first measure, with handling time tracked alongside it