Traditional music recommendation systems (RS) rely on collaborative filtering technique (CF) to recommend songs or artists in which recommendations are made based on the neighboring analysis of items/users. It is computationally efficient and performs well when the data is ideally full, when there are limited user inputs or few user/item inputs, it immediately lost its competitive advantage. Additionally, traditional RS techniques including content-based one heavily rely on explicit user feedback (e.g. user rating) to generate recommendations. In music/song recommendation, however, implicit feedbacks such as play frequency, play list prevail. Making recommendation on such implicit feedbacks requires efficient and accurate latent factor learning techniques to construct user or item feature space, which is inherently computationally costly. This paper presents a new and lightweight classification model for Chinese song RS based on computational analysis of the lingual part of song lyrics. Through extracting and combining the term frequency and inverse document frequency (tf*idf) from song lyrics, we construct a composite emotion point matrix for each song which can then be used to further classify songs based on its inherent emotion and make recommendation accordingly.