Everyday thousands of new information resources are linked to the web. This way the web is growing very fast what makes search tasks more difficult. To solve the problem several initiatives were undertaken and a new area of research and development emerged: the one called Semantic Web. When we refer to the semantic web we are thinking about a network of concepts. Each concept has a group of related resources and can be related to other resources, thought we can use this concept network to navigate among web resources or simply among information resources. From the undertaken initiatives one became an ISO standard: Topic Map ISO 13250. The aim of this paper is to introduce a Topic Map™ Builder, that is a processor that extracts topics and relations from instances of a family of XML documents. A Topic Map Builder is strongly dependent on the resources structure. So if we have an heterogeneous set of information resources we have to implement several Topic Map Builders. This is not very useful, and to overcome this problem we have created an XML abstraction layer for Topic Map Builders. To describe that process, i.e. the extraction of knowledge from XML documents to produce a TM, we present a language to specify topic maps for a class of XML documents, that we call XSTM (XML Specification for Topic Maps). We also discuss a XSL processor that automatically generates the Extractor from its formal specification, the XSTM-P.