In this work we investigate a data-driven vector representation of word embedding for the task of classifying song lyrics into their semantic topics. Previous research on topic classification of song lyrics has used traditional frequency based text representation. On the other hand, empirically driven word embedding has shown sensible performance improvment of text classification tasks, because of its ability to capture semantic relationship between words from big data. As averaging the word vectors from a short text is known to work reasonably well compared to the other comprehensive models utilizing their order, we adopt the averaged word vectors from the lyrics and user's interpretations about them, which are short in general, as the feature for this classification task. This simple approach showed promising classification accuracy of 57%. From this, we envision the potential of the data-driven approaches to creating features, such as the sequence of word vectors and doc2vec models, to improve the performance of the system.