Encoding and Ranking Similar Chinese Characters

摘要

Automatically detecting similar Chinese characters is useful in many areas, such as building intelligent authoring tools (e.g. automatic multiple choice question generation) in the area of computer assisted language learning. Previous work on the computation of Chinese character similarity focused on detecting character glyph similarity while ignored the importance of other character features, such as pronunciation and meaning. In this article, we present a way to encoding 4,500 simplified Chinese characters in terms of character glyph, pronunciation and meaning, annotating similar Chinese characters and automatically ranking similar characters based on the approach of learning to rank. The experiment results indicated that this approach could be useful for ranking and recognizing similar Chinese characters in terms of glyph, pinyin and semantic meaning. Moreover, it has been found that the learning to rank Listwise (ListNet) method was more effective than Pointwise (MART) and Pairwise (RankNet).