Open Bengali OCR Benchmarks Accelerate Archival Newspaper Digitization
University NLP groups release a clean evaluation corpus of historic print layouts, cutting character error rates on faded mid-century broadsheets.

SYLHET — Digitizing South Asia’s twentieth-century print heritage has long stumbled on a technical hurdle: off-the-shelf optical character recognition models falter on multi-column newspaper layouts, faded lead-type ink, and complex Bengali conjunct characters (juktakkhor).
A collaborative research group spanning computer science labs in Dhaka and Sylhet has published an open evaluation dataset containing 18,000 manually verified archival newspaper lines. Early open-weight vision-language fine-tunes evaluated against the benchmark reduced word error rates by nearly 40 percent compared with legacy pipelines.
ABOUT THE REPORTER
Nafisa Kabir
Technology & Engineering Reporter
Follows university robotics labs, agritech hardware, and public digital infrastructure.
Related Dispatches
More in Technology
Engineering Students Build Low-Cost Grain Moisture Telemetry Nodes
Field trials in Bogura and Dinajpur warehouses show open-hardware moisture alerts can help smallholder cooperatives prevent fungal spoilage during humid monsoon weeks.

University Research Clubs Expand Their National Collaboration
Student-led research organizations across Dhaka, Chattogram, Rajshahi, and Khulna are replacing isolated annual symposiums with a shared national working-paper network.

Instant Merchant Settlement Pilots Reshape Dhaka SME Liquidity
Commercial banks and mobile financial service providers are testing real-time clearing rails aimed at drawing cash-heavy wholesale markets into formal ledger histories.