PDF Annotation & Transcription Experts – Bengali

Mercor is seeking Bengali‑language specialists to build high‑quality training data for document‑understanding models. The position focuses on creating detailed structural maps and accurate transcriptions of publicly available Bengali PDFs, covering a wide range of real‑world formats such as newspapers, textbooks, exam papers, flyers and manuals. By enriching datasets for scripts that current vision‑language systems struggle with, contributors help improve AI performance on Bengali and other Indic scripts. Each workday involves locating a suitable Bengali PDF that contains at least one multimodal element—such as an image, table, diagram or handwritten section—and documenting its source. Contributors then delineate every meaningful region on the page, assign it a component type, and set a reading‑order index. They link each region to its parent figure or table, transcribe the text exactly as it appears—including handwritten portions—and flag any illegible areas. Additional metadata about language, document type, page dimensions and the presence of tables, formulas or handwriting is recorded. After completing a task, the annotation is reviewed by another Bengali expert, and seasoned annotators may also perform peer reviews. Applicants must be native Bengali speakers with complete mastery of the script, its diacritics and conjunct forms. Prior experience with document‑related work—such as annotation, transcription, translation, localization, subtitling, proofreading, journalism or regional‑language data review—is required. Candidates need to prioritize character‑level precision, apply a consistent taxonomy across many pages, and feel comfortable handling unfamiliar layouts like multi‑column newspapers or handwritten forms. Desirable backgrounds include work on regional‑language AI data, OCR correction, bilingual quality assurance, typesetting, copy‑editing or Unicode normalization. The project values accuracy above speed, with handling time monitored alongside quality metrics.