GRU achieves 74.03% accuracy in Dzongkha next word prediction

A new system has been developed to predict probable words for Dzongkha, aiming to reduce the number of keystrokes required for typing this complex script. The project utilizes a dataset of 100,000 sentences acquired from DCDD to train and evaluate three distinct models: LSTM, Bi-LSTM, and GRU.

• The dataset contains 1,331,282 words and 28,344 unique words, processed through tokenization, N-gram generation, and padding.
• Hyperparameters for all three models were fine-tuned to optimize performance on the Dzongkha language task.
• The GRU model outperformed the others with an accuracy of 74.03%, while also effectively handling overfitting due to its lightweight nature.

This system addresses the challenge of documenting and typing Dzongkha by providing a convenient tool that retains cultural value through efficient text prediction.