Improved word similarity computation for Chinese using sub-word information

In the Chinese language, words consist of characters each of which is composed of one or more components. Almost every individual Chinese character has a specific meaning, and the meaning of a word is usually highly related to the characters that comprise it. Likewise, sub-character components often make a predictable contribution to the meaning of a character, and in general characters that have the same components have similar or related meanings. It is easy to automatically decompose words into characters and their components. In this paper, we improve on a corpus-based approach to computing word similarity in Chinese by extending it according to the characters and components shared between words. In an evaluation on 30,000 word types (noun, verb and adjective), we obtain a 39% relative improvement compared with a state-of-the-art baseline.