Background
This project is part of of the Bengali.AI Handwritten Grapheme Classification Kaggle competition. The organizers of the competition provide the following motivation for the task:
Bengali is the 5th most spoken language in the world with hundreds of million of speakers. It’s the official language of Bangladesh and the second most spoken language in India. Considering the language’s reach, there is significant business and educational interest in developing AI that can optically recognize images of the language’s handwritten script. This challenge hopes to improve on existing approaches to Bengali handwriting recognition.
Optical character recognition is particularly challenging for Bengali. While the script has 49 letters (to be more specific 11 vowels and 38 consonants) in its alphabet, there are also 18 potential diacritics. This means that there are many more graphemes, or the smallest units in a written language. The added complexity results in ~13,000 different grapheme variations (compared to English’s 250 graphemic units).
A sample of the Bengali Alphabet from OhFact.com
Bangladesh-based non-profit Bengali.AI is focused on solving this problem. They build and release crowdsourced, metadata-rich datasets and open source them through research competitions. Through this work, Bengali.AI hopes to democratize and accelerate research in Bengali language technologies and to promote machine learning education.
For this competition, participants are given the image of a handwritten Bengali grapheme and are challenged to separately classify three constituent elements in the image: grapheme root, vowel diacritic, and consonant diacritic.
Through their innovation and hard work, particpants will accelerate Bengali handwritten optical character recognition research and help enable the digitalization of educational resources. Moreover, the methods introduced in the competition will also empower cousin languages in the Indian subcontinent.
Source Code and Data
Check out the GitHub repository which contains our source code and data!