Generic Plan Recognition for Dialogue Systems

Generic Plan Recognition for Dialogue Systems George Ferguson, James F. Allen University of Rochester Rochester, NY, 14627-0226 ABSTRACT We describe a general framework for encoding rich domain mod- els and sophisticated plan reasoning capabilities. The approach uses graph-basedreasoning to address a wide range of tasks that typically arise in dialogue systems. The graphical plan representation is inde- pendent of but connected to the underlying representation of action and time. We describe types of plan recognition that are needed, illustrating these with examples from dialogues collected as part of the TRAINS project. The algorithms for the tasks are presented, and issues in the formalization of the reasoning processes are discussed. 1. Introduction Plan recognition is an essential part of any dialogue system. Traditional approaches to plan recognition are inadequate in one of two ways. Those that are formally well-specified tend to be highly restricted in the phenomena they can accomo- date and are therefore unsuitable for a general purpose dia- logue system. On the other hand, the heuristically-motivated systems have been difficult to formalize and hence to under- stand. In both cases, the representation of plans is insufficient for a collaborative dialogue-based system. The research reported here is part of the TRAINS project [1]. The goal of this project is an intelligent planning assistant that is conversationally proficient in natural language. In this paper we concentrate on the plan recognition procedures of the domain plan reasoner component of the system. As examples of the phenomena that arise in discourse and af- fect plan recognition, consider the following utterances gath- ered from TRAINS dialogues: 1. Utterances that suggest courses of action, e.g., (a) Send engine E3 to Dansville. (b) Move the oranges to Avon and unload them. This is the prototypical case studied in the literature, and most systems are limited to handling only this case. 2. Utterances that identify relevant objects to use, e.g., (a) Let's use engine E3. (b) There's an OJ factory at Dansville. The second sentence is an example of an indirect sug- gestion to use the OJ factory. 3. Utterances that identify relevant constraints, e.g., (a) We must get the oranges there by 3 PM. (b) Engine E2 cannot pull more than 3 carloads at a time. 4. Utterances that identify relevant lines of inference, e.g., (a) The car will be there because is it attached to en- gine El. 5. Utterances that identify goals of the plan, e.g., (a) We have to make OJ. 6. Utterances that introduce complex relations, e.g., pur- pose clauses such as (a) Use E3 to pick up the car. (b) Send engine E3 to Dansville to pick up the oranges. Our approach to plan reasoning is motivated by examples such as these. It is a generic approach because the details of the algorithms do not depend directly on properties of the underlying knowledge representation. Rather, the approach assumes that certain operations are exported by the under- lying reasoner (such as entailment, ~), and it uses these to validate plan reasoning steps. …

Content truncated to 500 words