This paper addresses the problem of taking into account the feasibility and cost of verbal communication actions at the task planning level in the context of human-robot interaction. By determining the human and robot context in which the referring expressions to entities can be effectively achieved and at which cost, our planner can 1) prevent potential deadlocked situations where it is not possible to produce a feasible communication act 2) find plans reducing the overall communication complexity 3) evaluate different communications strategies and select the most suitable one. Our approach is based on the extension of a multi agent hierarchical task planner capable of maintaining, throughout the planning process, an estimate of the human partner knowledge about the environment using a semantic representation. A Referring Expression Generation (REG) can then be executed for each communication action allowing the task planner to be informed about its feasibility and cost during the plan elaboration process. Three scenarios are presented to validate the method, where a robot drives the human actions to perform a cube arrangement task.