Hierarchical Joint Learning for Chinese Word Embeddings
Zhe Wang, Qiong Wang, Yechao Bai, Xinggan Zhang · 2019 IEEE International Conference on Signal, Information and Data Processing (ICSIDP) · 2019
In recent years, some methods of learning Chinese word embedding based on the characters and sub-characters information have been proposed and achieve state-of-art results. Generally speaking, good embedding of characters and sub-characters can help learn better embedding of words. However, current methods utilizing characters and sub-characters information, such as character-based CWE and sub-character-based JWE, only learn embedding of characters and sub-characters by predicting the target word. We believe it's not enough to improve the quality of characters and sub-characters embedding, so we propose a method called HJWE which predicts the target word, characters and sub-characters in the target word at the same time. The experiment result shows that our method performs best on the word similarity, word analogy and text classification tasks.