Combining Key-target Resizing and LSTM Language Model to Reduce Typos in Thai Soft Keyboard

Abstract

Recently virtual keyboard has become one of the main user interfaces for entering textual data to a smartphone. For virtual keyboards in foreign languages, there are many researches that study how to reduce typos caused by the small size of each button in the virtual keyboard. Nevertheless, as we do not find this kind of researches for Thai virtual keyboard, we propose our work that experiments and evaluates feasibility of using a combination of language model and key-target resizing technique to reduce typos on Thai virtual keyboard. Our work starts by using standard Thai vocabulary corpuses to train two language models (i.e., Markov Chain and LSTM) in order to predict the most likely buttons that a user will press next. Then, we collect typing data on Thai virtual keyboard from seven users using our prototype system. Finally, we analyze the collected data in conjunction with predicted results from our language model. According to our experimental results, the LSTM based language model performs better than the Markov Chain based language model in predicting the next Thai’s character buttons. When this LSTM language model is used to enlarge six buttons with highest predicted probabilities in advance, results show that it helps reduce typos by 5.05%. More specifically, the number of typos is reduced by 13 out of 257 typos.