Word Completion and Sequence Prediction in Bangla Language Using Trie and a Hybrid Approach of Sequential LSTM and N-gram
Soumik Sarker, Md. Ekramul Islam, Jillur Rahman Saurav, Md Mahadi Hasan Nahid · 2020
Autocompletion and sequence prediction is the basis of any assistance systems. When we step out to type something, it is much comforting to get a suggestion of the next word or even the full sentence before we type, which saves keystrokes of typing and reduce misspelling. Despite the several promising works in English language, little prior research in Bangla has shed light on this domain. In this paper, we proposed an integrated methodology of trie, sequential LSTM and N-gram for word completion and sequence prediction in Bangla language. The trie data structure was implemented to store Bangla vocabulary and retrieve the word from user-inputted prefix. For sequence prediction, we explored a hybrid approach of neural network and N-gram. This collaboration of sequential LSTM and N-gram reveals a better performance than any single model implementation. We evaluated this model with both small and large-scale Bangla datasets for better efficiency. The experiments show a promising outcome of our hybrid approach for word completion and sequence prediction. We believe that our framework leads to a profound impact on Bangla search engines, keyboards, and further researches based on recommendation systems.