Knowledge Acquisition for Content Selection

An important part of building a natural-language generation (NLG) system is knowledge acquisition, that is deciding on the specific schemas, plans, grammar rules, and so forth that should be used in the NLG system. We discuss some experiments we have performed with KA for content-selection rules, in the context of building an NLG system which generates health-related material. These experiments suggest that it is useful to supplement corpus analysis with KA techniques developed for building expert systems, such as structured group discussions and think-aloud protocols. They also raise the point that KA issues may influence architectural design issues, in particular the decision on whether a planning approach is used for content selection. We suspect that in some cases, KA may be easier if other constructive expert-system techniques (such as production rules, or case-based reasoning) are used to determine the content of a generated text.