Pragmatic knowledge for resolving ill-formedness
Lance Ramshaw, Ralph Weischedel · 1989
Many examples of ill-formedness that people routinely and easily correct can be resolved by a natural language system only if it makes use of knowledge of the pragmatic context. This investigation centers around examples of alias errors, where the ill-formedness is due to a single word that is incorrect but still lexically understood, as with the substitution of on for in in the phrase stay on good shape. Localizing and resolving such errors frequently depends on pragmatic knowledge. This thesis presents a model for pragmatic context within expert advising dialogues, where an agent who is building a plan to solve a problem consults with a domain expert, and develops methods for applying that model to resolving ill-formed input. Metaplans are used to model the structure of the agent's problem-solving behavior, both the gradual refinement of the domain plans being considered and the connection between them and the queries motivated by them. The partially-specified domain plans that the agent is considering are represented by nodes in a plan classification hierarchy, and these classes of domain plans in turn serve as arguments to the problem-solving metaplans. The expansion and search of the metaplan tree that models the problem-solving context is governed by heuristics based both on its metaplan structure and on a model of the agent's world knowledge. This model can be used to track the problem-solving moves implicit in a sequence of well-formed queries and also to predict likely moves and queries as determined by the context which can then be linked to the partial interpretation of an ill-formed query suggesting corrections for the ill-formedness. This approach has been implemented in a system called Pragma, which suggests corrections based on pragmatic context for alias errors in naval domain queries, using techniques that could also be extended to other classes of ill-formedness and to generating cooperative responses. Pragma demonstrates that a model capturing the pragmatic structure of a particular discourse setting can be used to increase the robustness of a natural language interface.