A Description Logic for NLP

In the late eighties inference in KL-ONE was shown to be undecidable. Since then the emphasis in research has been on developing and investigating systems that are computationally well behaved, i.e. are tractable or at least decidable. As a result many commonly used description logics (also known as terminological logics or KL-ONE-based knowledge representation formalisms) have restricted expressiveness and are in their current form not suitable for natural language applications. This is evident, for example, from Schmidt [14] who links knowledge representation with a relational approach to natural language semantics. For encoding knowledge formulated in a very limited fragment of English we already need the full expressive power of role constructs which have been eliminated in many languages.

We share the view of Doyle and Patil [4] who argue for expressiveness as opposed to computational efficiency. Our experience with users interested in user modelling and natural language simulations can be summarized as follows:

1. Users want expressiveness.
2. They want representation languages with more basic features than just concepts, roles and individuals (i.e. A-Box elements) and operations on these.
3. And, they want special inference tools.

1. They communicate in natural language.
2. They actively pursue complex goals, which may be conflicting.
3. They have the means of analyzing (some of) the pragmatic content of what is being said, i.e., they have a deeper understanding of `belief', `intension' or `argument'.

In the following sections we describe MOTEL and the different extensions we are working on.