Inner-character and Inner-word Features Based Representation Learning for Chinese Word Embedding

Chinese word embedding is a significant task in natural language processing (NLP). Most researchers explored Chinese word embedding according to radical, component, stroke n -gram and character features. Besides these features, Chinese characters still have structure and pinyin characteristics. In this article, we propose ensemble ssp2vec and connective ssp2vec to utilize inner-character features (stroke, structure, and pinyin) for learning Chinese word embeddings. Then we design hierarchical ssp2vec to forecast the contexts according to the combination of inner-character (stroke, structure, and pinyin) and inner-word features (character) of Chinese words to explore different feature combination ways for learning feature relevance and comprehending word semantics, where feature substring is proposed to learn the relevancy of stroke, structure, and pinyin. Experimental results for word analogy, word similarity, text classification, and named entity recognition tasks demonstrate that the proposed methods outperform most state-of-the-art models.