We present a novel scheme of spoken dialogue systems which uses the up-to-date information on the web. The scheme is based on information extraction which is defined by the predicate-argument (P-A) structure and realized by semantic parsing. Based on the information structure, the dialogue system can perform question answering and also proactive information presentation. Feasibility of this scheme is demonstrated with experiments using a domain of baseball news. In order to automatically select useful domain-dependent P-A templates, statistical measures are introduced, resulting to a completely unsupervised learning of the information structure given a corpus. Similarity measures of P-A structures are also introduced to select relevant information. An experimental evaluation shows that the proposed system can make more relevant responses compared with the conventional "bag-of-words" scheme.