A program that turns handwritten Bangla into typed text won the first "AI for Bangla" competition on Tuesday, 22 February 2022. Team Obhijatrik took the Tk 1 lakh first prize at the BCC auditorium in Agargaon's ICT Tower, the day after International Mother Language Day, for a project that does something every bank, school and land office in the country still pays clerks to do by hand.
The winners
- First, Tk 1 lakh: Obhijatrik, Bangla handwriting to printed text (optical character recognition for handwriting).
- Second, Tk 50,000: SUST Dreamers, an "NLP Bangla Toolkit", a library of building blocks for processing Bangla text.
- Joint third, Tk 25,000 each: BUET Anumiti, for a natural-language-inference dataset, and Intelligent Machines Ltd, for reading handwritten Bangla on receipts.
The contest drew 161 entries from young researchers and developers. A jury led by writer and academic Muhammed Zafar Iqbal picked a top ten that also included XML Indic (modelling Indian languages), Beyond the Hills (translation tools for Bangladesh's small ethnic languages), Ekush (a Bangla handwriting dataset), DAB (authorship attribution), Roktim (a named-entity dataset) and Innovation Garage (a document reader for blind users). All ten received honourable mentions.
Why datasets are the real prize
Three of the seven awarded projects are not products but datasets and toolkits. That is not a consolation prize. English has decades of labelled text and speech to train models on; Bangla, the world's sixth most spoken language, has comparatively little that is open and clean. A handwriting dataset such as Ekush, or an inference dataset from BUET, is what lets the next hundred teams build something without starting from zero. It is also why the organiser was the Bangladesh Computer Council's "Enhancement of Bangla Language in ICT through Research and Development" project, whose remit is exactly that groundwork.
What was said
"Research on Bangla is happening all over the world today," Zafar Iqbal told the winners. "Those who are presenting Bangla to the world through technology are today's language soldiers." ICT Division senior secretary N M Zeaul Alam said the competition would be held every year. BCC research director Ehsanul Parvez and North South University's Nabil Muhammad also spoke.
For anyone who has typed out a handwritten prescription, a land deed or a school register letter by letter, the winning entry is the first sign that the machine may soon do it instead.
