Renovating word vectors to build Chinese sentiment lexicon

Sentiment lexicon is the core of many academic and commercial sentiment analysis system. But compared with the plentiful English sentiment resources, Chinese sentiment lexicon is scarce which limits the application of lexicon-based method in Chinese sentiment analysis. In this work, we proposed a novel architecture to produce Chinese sentiment lexicon. At first, we trained word vectors from distributional information of words in large corpora. The capacity of these word vectors in sentiment analysis is confined because of the noisy information in their feature space. So a feature selection method was used to renovate the word vectors. After that, the renovated word vectors was used with a similarity-based method to produce the Chinese sentiment lexicon. To evaluate the usefulness of our lexicon, both qualitative and quantitative experiments were designed. The results show that it outperforms previously studied lexicons and indicate that our sentiment lexicon could be used as an important resource for sentiment classification tasks.