Another important question is how to represent knowledge so that it can be viewed from different perspectives or viewpoints. The latter is particularly important for NLP systems, since context, user characteristics and goals often highlight' some objects and attributes, i.e., they become more salient than others. For instance, [McCoy 89] demonstrates how perspectives can be used to generate context-sensitive responses that correct user misconceptions. The system relies on a domain ontology and a set of perspectives which assign salience values to object attributes. Therefore, only relevant objects and attributes can be included in the explanations depending on the current perspective (e.g., a house can be viewed as a place to live or a building). Object similarity is also computed as a function of overlapping and distinct attributes. However, all perspectives must be encoded in advance and related to user goals if required. This might be achievable for expert/database systems where the knowledge engineer can forsee all modes of interaction but in explanation and advice-giving systems a more general approach is required.