Hybrid model for classifying Indo Aryan and Tamil texts from historic manuscripts

The ancient manuscripts, especially Indo-Aryan and Tamil texts have the complex linguistic structure in their manuscripts with historical differences. The present paper is a BERT-Li Scribe Hybrid Model of the historic Indo-Aryan and Tamil manuscript classification. The model combines the contextual embedding of BERT, which learns the semantic links in the text, and LiScribe, a specialized sequence model that learns features, and linguistic patterns of Indo-Aryan and Tamil scripts at a character level. The sample of 1,055 manuscripts of the Library of Congress and University of Hamburg is a perfect combination of Indo-Aryan and Tamil texts. The model that is offered provides the classification of two large groups, such as Indo-Aryan (Hindi, Bengali, Marathi, Gujarati) and Tamil, which guarantees the level of classification on a family level and specific scripts. Training is performed using categorical cross-entropy loss and Adam optimizer with learning rate scheduling with dropout layers used to avoid overfitting with noisy historical data. The model, which was coded in Python and deployed with the help of such libraries as TensorFlow and PyTorch, demonstrated a high overall classification accuracy of 97.61%, being able to distinguish between the Indo-Aryan and Tamil texts at the same time. The attention mechanism also increases the concentration on the important features by the model even in the degraded manuscripts. This mixed methodology proves the usefulness of the combination of deep learning and linguistic feature extraction to the correct classification of historical manuscripts.