Natural language generation from plans

TLDR

Plan structures produced by AI planners often lack the explanatory detail and natural semantics needed for fluent natural‑language descriptions, a challenge similar to that faced by expert systems and due to planners’ limited world models. This paper aims to design a system that takes a plan structure from AI planners and produces natural‑language explanations of how to execute the plan. The system maps the hierarchical plan structure to natural‑language text, leveraging the plan’s organization to guide the generation. Experiments with the NONLIN planner show that the system can produce promising but still rough natural‑language explanations, with primitive planner constructs often rendered clumsily and the range of language constructs limited by the planner’s shallow representations.

Abstract

This paper addresses the problem of designing a system that accepts a plan structure of the sort generated by AI planning programs and produces natural language text explaining how to execute the plan. We describe a system that generates text from plans produced by the NONLIN planner (Tate 1976).The results of our system are promising, but the texts still lack much of the smoothness of human-generated text. This is partly because, although the domain of plans seems a priori to provide rich structure that a natural language generator can use, in practice a plan that is generated without the production of explanations in mind rarely contains the kinds of information that would yield an interesting natural language account. For instance, the hierarchical organization assigned to a plan is liable to reflect more a programmer's approach to generating a class of plans efficiently than the way that a human would naturally chunk the relevant actions. Such problems are, of course, similar to those that Swartout (1983) encountered with expert systems. In addition, AI planners have a restricted view of the world that is hard to match up with the normal semantics of natural language expressions. Thus constructs that are primitive to the planner may be only clumsily or misleadingly expressed in natural language, and the range of possible natural language constructs may be artificially limited by the shallowness of the planner's representations.