Sentiment analysis draws increasing attention of researchers in wide-ranging fields. Compared with the commonly-used categorical
approach representing affective states as a few discrete classes, the dimensional approach represents emotions as continuous numerical values in multiple dimensions, such as valence-arousal (VA) space. It can thus provide more fine-grained sentiment analysis. However, affective lexicons with VA ratings are very rare. This limitation makes the dimensional approach hard to use in reality. This study proposes a VA ratings prediction method combining word2vec and KNN.