Multi-task Learning and Thai Handwritten Text Recognition

Written languages have some commonalities and differences. Knowing one language can help learning other languages better. While labelled data is limited in offline Thai handwritten text recognition problem domain, multi-task learning is chosen to address the problem in our study. The multi-task model is trained from Thai, Latin and Devanagari writing scripts. The model consists of three layers employing Convolutional Neural Network, Recurrent Neural Network and Connectionist Temporal Classification, respectively. Recognition accuracies are compared against three corresponding single-task models. Thai handwritten text recognition performance is considerably improved by multi-task learning. Learning multiple languages helps the model generalize better when trained by large enough datasets.