Noisy language modeling framework using neural network techniques

The text entry interaction between human and computer could be noisy. For example, the typing stream is a reflection of user typing behaviours which include user particular vocabulary, typing habits and typing performance. As computer users inevitably make errors, a typing stream generated from using computer QWERTY keyboard implies all users' self-rectification actions rather that a clean text. Therefore this research develops a novel intermediate layer language modeling framework called ALMIL (i. e. Adaptive Language Modelling Intermediate Layer) which is seen as a communication language layer between human and computer to analyze noisy language stream and provide users with two fundamental functions, namely Text Prediction and Text Correction. A specific research case of ALMIL called Intelligent Keyboard (IK) aiming to develop a user oriented hybrid framework with self-adaptive function to help people using QWERTY keyboard more effectively is also conducted. In order to explore the methodologies, influential factors and demonstrate the feasibility of the frameworks, a comprehensive neural networks and language modeling process is carried out. Several neural network models which include a Focused Time-Delay Neural Network model (FTDNN) to model non-noisy, noisy and typing stream datasets, a Time Gap Neural Network model (TGNN) to simulate and predict user typing time gap between two consecutive letters, a Prediction using Time Gap model (PTG) to predict right symbols based on user typing speed, a Probabilistic Neural Network based model (PNN) to simulate 'Hitting Adjacent Key Effors', and a Word List real-time ranking model (VvLR) on prioritizing prediction results are developed. All the models have been demonstrated, and shown high performance through a set of experiments using a range of dataset. In essence, this research brings forth a creative concept - intermediate layer language modeling framework for noisy language processing, pioneers a comprehensive neural networks modelling process, and originally develops a hybrid solution to combine multiple correction functions based on an evolutionary ranking approach. It produces a significant contribution in the area of neural networks application and shows a direction for Human-Computer noisy language interaction research. Also a full report on disabled people typing behaviour, a development of EIM application and a universal pre-processing tool for all neural networks modelling and n-gram, calculation will benefit both research and commerce.