The Input And Predictive Performance Optimization Of Chinese Input IME On Touch Screen

With the development of science and technology in recent years,smart phones become increasingly popular,and human-computer interaction becomes more and more complex.Input method engine(IME),as an important text input software,has its irreplaceable position.In the field of Chinese IME,the Chinese Pinyin IME is undoubtedly the most popular input method.This thesis focuses on improving the performance of Pinyin IME in terms of touch-screen keyboard,language model,and the overall structure.Compared to the physical keyboard,the virtual keyboard of the touch-screen device does not have the tactile sense and the percussion feeling of the keys.In addition,the different touchscreen devices have different screen sizes,which makes it hard to determine the size of keyboard layouts.Instead of designing a ”general” keyboard for all screens,it's better to get the keyboard itself to fit the screen and the user's input habits.With the reinforcement learning method,we designed an adaptive keyboard:When the user presses a key,next ”possible” keys will become bigger and move closer to the current position of finger.Besides,the smart keyboard will learn the most suitable keyboard based on the input speed of the user.We analyzed the typical work flow of modern Pinyin IME.Basically the core engine of IME is a pipeline of three parts: pinyin segmentation,candidate words fetching and candidate sentence generation.In this framework,the language model is regarded as the core of Pinyin IME.Neural network language models(NNLMs)have been shown to outperform traditional ngram language model.However,too high computational cost of NNLMs becomes the main obstacle of directly integrating it into pinyin IME that normally requires a real-time response.In this thesis,an efficient solution is proposed by converting NNLMs into back-off n-gram language models(BNLMs),and we integrate the converted NNLM into pinyin IME.Our experimental results show that the proposed method gives better decoding predictive performance for pinyin IME with satisfied efficiency.Previous studies and experiments are based on the typical architecture of modern pinyin IME,how to break the shackles of traditional workflow has become an important issue we are considering.As two typical applications of natural language processing,input method and machine translation have been regarded as two different research fields.However,we try to solve the problem of input method with the idea of machine translation,espectially Neural Machine Translation(NMT),with a strong sense of inspiration.