Recently, the government of South Korea has offered a variety of welfare programs that are customized to diverse demands, such as diabetes management, alcohol addiction rehabilitation, living condition improvement, etc. These welfare programs have become too diverse to be remembered and recommended by individuals, and the government now has a list matching program recipients and programs for further studies. This research investigates such welfare program recommendation with a conditional variational autoencoder merged with collaborative filtering, a.k.a. CVAE-CF. We use a natural language description to provide the program information, or item in the context, and we utilize the demographic information from potential recipients as the user information. Our results show agreeable performance for future application to recommendation tasks showing 63% recall and 13.1% precision on average.