A Deep Reinforcement Learning-based Chinese Pinyin IME

Input Method Editor(IME)is an popular topic among human-computer interaction territory.The main responsibility of an IME is to help users enter characters or symbols that cannot be entered with standard input peripherals into a target device directly.It typically involves the interactions that user makes with the input peripheral and converts these information into the corresponding character and symbol sequence to be entered into the target device.Chinese pinyin IME was first introduced in the 1970 s,and has been evolved in technology and optimized in methodology since then.Now Chinese pinyin IMEs typically support sequence-to-sequence conversion,and may also auto-completion,typo-correction and some other helpful functions,which makes them become the most popular Chinese IME.Touchscreen devices,especially touchscreen mobile phones become popular recently.IMEs for touchscreen devices generally follow the keyboard-based input style,which usually display a virtual keyboard for the user to input alphabets by touching the virtual keys in it.Comparing to the traditional keyboards,however,the virtual keyboards on touchscreen devices can collect richer details of the user finger gestures,including the movement velocity,touchpoint size,touch pressure,etc.These information can be taken into consideration for understanding the user intention,while still few research have been conducted.On the other hand,with the development in machine learning,especially that in deep learning and neural networks,the application of deep learning algorithms in natural language processing tasks,for example,machine translation,dialog generation,part-of-speech tagging has received significant success.Chinese pinyin IMEs are mainly equipped with these algorithms to optimizing the language model behind to improve the pinyin-to-character conversion accuracy.This work intends to improve the user experience of Chinese pinyin IMEs.On the on hand,this IME allows user finger sliding as input and converts the sliding into the target pinyin sequence.Moreover,the layout of the virtual keyboard will also adapt during the sliding process to help users input more efficiently.Most of the parameters for this virtual keyboard are learned with a deep reinforcement learning algorithm other than defined manually.A seq2 seq model is also utilized to convert the pinyin sequence to the corresponding Chinese sequence,so that the final target Chinese sequence can be entered.Experiments showed that this IME can improve the input efficiency comparing with baseline ones.The key advantage of this IME is that the rules utilized are learned automatically from user behavior other than defined by human so less labor is required when building it.Such an idea may also be helpful for some other human-computer interaction tasks.