A Word Representation Method Based on Glyph of Chinese Character

摘要

The quality of words representation has an important impact on natural language processing tasks. Aiming at the problems in the current Chinese word representation method: the training data set is huge, the model quality depends on the data set, and the model stability is poor, a word representation method based on the glyph of Chinese character, Glyph2Vec, is proposed. Taking full advantage of the semantic information contained in Chinese characters, a glyph auto-encoder is constructed based on a convolutional auto-encoder. The glyph auto-encoder is used to obtain Chinese character embedding by mapping the glyph of Chinese character in the potential low-dimensional semantic space. In the Chinese named entity recognition task experiment, Glyph2Vec improves the accuracy to F1 score by 0.77%, 1.84%, and 1.31% respectively, compared with Word2Vec. The experimental results show that the method proposed is better than the existing results, which proves the effectiveness of this method.