As a result of the growth of online music streaming services, a large number of playlists have been created by users and service providers. The title of each playlist provides useful information, such as the theme and listening context, of the songs in the playlist. In this paper, we investigate how to exploit the words extracted from playlist titles for text-based music retrieval. The main idea is to represent songs and words in a common latent space so that the music retrieval is converted to the problem of selecting songs that are the nearest neighbors of the query word in the latent space. Specifically, an unsupervised learning method is proposed to generate a latent representation of songs and words, where the learning objects are the co-occurring songs and words in playlist titles. Five metrics (precision, recall, coherence, diversity, and popularity) are considered for performance evaluation of the proposed method. Qualitative results demonstrate that our method is able to capture the semantic meaning of songs and words, owning to the proximity property of related songs and words in the latent space.