The word-level sentiment analysis is an essential issue in opinion mining. One challenge in this field is that not so many lexical items as expected have been labeled with sentimental opinions, especially in Chinese. There are two ways of rating words: one is manual marking which costs lots of resources, energy and time; the other is machine marking which is efficient, convenient and time-saving. There are a few machine rating approaches such as linear regression, support vector regression and weighted graph method. This paper compares the three approaches of linear regression, vector regression and kernel models based on valence-arousal (VA) space in order to study an effective and accurate machine learning algorithm for increasing more Chinese affective words on the existing affective lexicon.