Despite the advance in automatic term acquisition (ATA) research, the use of ATA in terminology work is limited. One of the main reasons is that most ATA research focuses on automatic term extraction (ATE) from textual corpora and does not consider terminologies. This paper introduces automatic term generation (ATG), a new method of ATA, which explores existing terminologies, in which timeline information is also considered. We propose a new way of representing the structure of a terminology as a graph, which consists of nodes that represent constituent elements of terms and two types of links that represent intra-term dependency between constituent elements and term closure between the first and the last constituent element of terms. ATG is then defined as a set of link prediction tasks, i.e. predicting two types of links. Cliques of the graph represent the new generated terms. They are determined in a backward fashion by first selecting nodes for which term closure links are predicted and then by walking backward to term nodes for which intra-term dependency links exist. Experiments showed that our approach gives high performance in predicting new terms.