In natural language processing, incremental semantic composition is one of the most prominent issues. In the past, numerous approaches have been developed for assigning meaning to noun and verb phrases and their complements and modifiers. However, their inferential power is often too weak to be applied to practical applications, or the expressiveness of the representation language is so complex, that it leads to intractable inference procedures. As an answer to these problems, we have developed an approach that relies on Description Logics (DL) for handling semantic construction. First, we will discuss this appraoch and show how a semantic knowledge base can be setup dependant on EUROWORDNET 1 (EWN) as a linguistic ontology. Subsequently we will outline our experience with and demands on EWN. 1 What is EMBASSI and its Objective? EMBASSI (”Elektronische Multimediale Bedienund Service-Assistenz”) has been http://www.illc.uva.nl/EuroWordNet/ a German joint project sponsored by the German Fed. Ministery of Research2. Our contribution to this project consists mainly of three components: the dialogue manager, formal ontologies for several multilingual application domains, and the language generation component to communicate system utterances to the user. The long-term goal of our research is to design and implement a generic dialogue system for rational (spoken) dialogues that helps a user to achieve certain goals in terms of operations of a technical application system – e.g. an information system, a system for controlling devices, or any other kind of problem solving systems. One of its design criteria is the ability to recognize users’ intentions in order to establish corresponding subgoals and control their processing. Furthermore, it should enable mixed-initiative, flexible and cooperative conversations, provide a high level of robustness as well as scalability at the linguistic and application dimensions, and easy portability to new domains. In addition, it should be possible to integrate multilingual linguistic interaction with multimodal It aims to provide easy access for everybody to complex technical systems (A/V home theatre, car devices, and public terminals), encouraging multimodal as well as multilingual user input. forms of input and output such as graphical user interfaces, and – by means of appropriate devices – the recognition of deictic actions. 2 DL Models of Applications Applications are characterized by a DL terminology which models the concepts used for making propositions about application situations. Basically, EMBASSI’s knowledge base is composed of two parts: the EWN ontology, which encodes the linguistic meaning of words determined on an empirical basis, and the STANDARD UPPER ONTOLOGY (SUMO) (NP01), which is used as a generic base model for concepts of the application domain (see (Lu02)). 3 Semantic Construction This section discusses the issue of semantic construction during analyzing natural language input. We are using an incremental approach to the composition of semantic representations. The backbone of our approach is -DRT (Fis96). The parser builds Discourse Representation Structures (DRSes) (KaR93) incrementally and maps them onto ABoxes3 (see (BLG02)). The main question here is how the mapping of domain independent in terms of EWN to application specific language usage in terms of a domain model is done. In the discourse domain, referents usually refer to instances in the application domain. Such pairs of a discourse referent and a corresponding inA general characteristic of DL-Systems is that the knowledge base is made up of two components: the intensional one, called TBox, and the extensional one, called ABox. TBox is a general schema characterizing the classes of individuals to be represented, their general properties and mutual relationships, while ABox is a partial instantiation of this schema, containing assertions relating either individuals to classes, or individuals to each other. So given a concept language L, an ABox-statement in L has one of the forms (DL96): C(a) Concept Membership Assertion R(a, b) Role Membership Assertion where C is an L-concept, R is an L-role, and a, b are individuals. stance are represented by means of a special role called has-lex. For instance, in the definition AvEvent has-lex Program1 it is claimed that an AvEvent4 is related to a discourse referent of Program1. Consequently, all words that are assigned Program1 as a meaning in EWN, designate an instance of AvEvent in the application domain. The DRS: