Sparsity Regularization for Chinese Spelling Check

Anonymous

Abstract: The Chinese Spelling Check (CSC) research objective is to detect and correct the spelling errors in the input. Generally, the number of incorrect characters in the input is far less than the correct, so the error probability sequence of input sentence predicted by the detection module should be sparse and sharp. However, all existing work has ignored this problem. In this paper, we add a sparsity regularization item to the objective function to make the output of the detection module close to sparse and sharp. We study two kinds of regularization: L1 regularization and minimum entropy regularization. Extensive experiments on the SIGHAN show that the sparsity regularization proposed in this paper can effectively improve the performance of the CSC model while without increasing the computational complexity. In addition, the robustness experiment results show that our method is robust.