Generating effective natural language instructions based on agent expertise

Juliet C Bourne, University of Pennsylvania


The automatic production of Natural Language instructions, i.e. those suitable for use by humans, has become an active area of research in recent years. In order for computer-generated instructions to be useful, they must be effective in accurately conveying the actions that are to be carried out by an agent. Conveying termination information, or when to stop performing an action, has been the focus of my dissertation research as it is an important part of generating effective instructions. I have done a corpus analysis of 3000 simple step-by-step maintenance instructions to study how termination information is expressed in naturally occurring texts. Using insights gained from the corpus analysis as well as the simulation of virtual agents carrying out similar tasks, I have specified an action representation, rules for reasoning about action information (particularly termination information), and some differences between agents with different levels of expertise. SPUD, a Natural Language generator developed at the University of Pennsylvania, takes this information, as well as information about the syntactic constructions used in instructions, and reasons about the best way to convey action information in effective instructions for particular agents.

Subject Area

Computer science|Artificial intelligence

Recommended Citation

Bourne, Juliet C, "Generating effective natural language instructions based on agent expertise" (1999). Dissertations available from ProQuest. AAI9937702.