The paper discusses the construction principles and use of a semantic network within the dialogue system for continuous speech input, called SPICOS II. One of the main problems in continuous speech recognition is the very close coupling of syntactic and semantic constraints with the recognition process while maintaining their descriptive independence. The paper will show how the evaluation of the semantic network, that actually models the behaviour of words and their semantically acceptable relations, can be integrated into the linguistic search process and help to reduce combinatorial complexity. The same network is used to support the resolution of anaphoric expressions in various ways.