Robust Handwriting Post Processing Method Based On Integration Of Multi Language Model

With the rapid development of smart mobile devices, it is necessary to develop a input method with the ability of continuous handwriting. In this paper, a new pattern of handwriting input method has been presented after fully analysising the user's requirement. Based on this mode, a custom input keyboard on i OS platform has been developed, thus user can place much focus on writing itself, meanwhile, the system will automatically segment, recognize and improve the fully system performance by statistical language model.As Chinese characters have complex structure, changeable fo nt and different writing style, only rely on the single character recognition is not enough to improve the recognize performance further more. Using statistical language model to combine the context information and single character confidence is a valid method. But the traditional language model are always based on the N-gram model, despite added some external features to reduce the model confusion, it still can not get rid of the root causes from the N-gram model. This paper proposed two ways of solving this problem, one is use the neural network bidirectional heuristic and Skip-gram model for reference, raise 6 methods to enhance the robustness of the system, meanwhile using Recurrent Neural Network to enhance the best jump statistical model to find the optimal path.In this experiment, People Daily Corpus, Wiki Chinese Corpus, and Sogou online news corpus are used as the robust language model training data set, and test this result by part of CASIA-OLHWDB 2.0,2.1 and 2.2 continuous handwriting corpus(about 135 million characters online handwriting corpus). The result shows that the jump language model in this paper has make the first candidate accuracy increase 3.35%, meanwhile the modify F1 value has a great increase, it proved that the method is effected in solving the error transfer problem.