Learning Radicals From Tangut Characters

Deep learning has been successively applied in Tangut character recognition, the most expensive and time-consuming part of which is the data annotation. Each Tangut character consists of one to several radicals, i.e. graphical components, and the number of radicals is far less than the Tangut characters, therefore we could annotate the Tangut characters with their radicals rather than annotate each characters. We have applied multi-label learning in detecting the radicals contained in Tangut characters, based on which any Tangut character could be represented with its radical set. The multi-label classifier is trained on the character-level annotated data together with the "Tangut character-radical dictionary" that help generate the radical vector for each Tangut character. We use a similar architecture of deep neural network with the one for recognizing Tangut characters, and replace the softmax activation function of the output layer with sigmoid, and use the cross entropy loss function. The radical extraction neural network has been used in accelerating the Tangut character annotation. Besides Tangut characters, the method could be applied in recognizing other characters that have no annotated dataset.