A Chinese Text Classification Model Based on Radicals and Character Distinctions

Chinese characters are generally correlated with their semantic meanings, and the structure of radicals, in particular, can be a clear indication of how characters are related to each other. In the Chinese characters simplification movement, some different traditional characters have been transferred into one simplified character (many-to-one mapping), resulting in the phenomenon of 'one simplified character corresponding to many traditional characters. Compared to the simplified characters, the traditional characters contain richer structural information, which is also more meaningful to semantic understanding. Traditional approaches of text modelling often overlook the structural content of Chinese characters and the role of human cognitive behaviour in the process of text comprehension. Hence, we propose a Chinese text classification model derived from the construction methods and evolution of Chinese characters. The model consists of two branches: the simplified and the traditional, with an attention module based on the radical classification in each branch. Specifically, we first develop a sequential modelling structure to obtain sequence information of Chinese texts. Afterwards, an associated word module using the part head as a medium is designed to filter out keywords with high semantic differentiation among the auxiliary units. An attention module is then implemented to balance the importance of each keyword in a particular context. Our proposed method is conducted on three datasets to demonstrate validity and plausibility.