Efficiency Tradeoffs for Language and Action in Collaborative Tasks

The view that communication is a type of action has provided a basis for research in natural language processing for many years. However, models of task-oriented dialogic interaction fail to fully exploit the language-as-action perspective because there is no deep understanding of the interaction between the definition of the task, the autonomy of the agents doing the task, and efficient language behavior. In this paper, I examine the interaction of three factors:(1) agents’ autonomy; (2) agents’ resource bounds; and (3) the distribution task-relevant information. I motivate a number of hypotheses with examples from naturally occurring task-oriented dialogues where agents’ autonomy varies. Then I show that dialogue performance of two autonomous agents is better than the dialogue performance of one autonomous and one non-autonomous agent, independent of the information distribution, as long as the agents are resource bounded.