When subjects are required to describe spatial networks with a constraint on the starting point of their description, dominant linearization strategies can be evidenced in discourse. These strategies consist in producing descriptive sequences which minimize the number of elements stored in working memory and/or the duration of the storage. Analysis of description latencies reveals the effect of the cognitive load associated with network complexity and suggests that the implementation of optimal strategies depends on a more extensive analysis of networks by the subjects. Lastly, we report the first steps in the design of a text generator that simulates different kinds of natural descriptive strategies.