Symbol graph based discriminative training and rescoring for improved math symbol recognition

摘要: In the symbol recognition stage of online handwritten math expression recognition, one-pass dynamic programming algorithm can produce high-quality graphs in addition best recognized hypotheses. this paper, we exploit rich hypotheses embedded a graph to discriminatively train exponential weights different model likelihoods and insertion penalty. The training is investigated two criteria: maximum mutual information (MMI) minimum error (MSE). After discriminative training, trigram-based rescoring performed post-processing stage. Experimental results finally show 97% accuracy on test set 2,574 written expressions with 43,300 symbols, significant improvement obtained.