The main goal of this work is to present an automated method for building domain-specific corpus-based ontologies. The assumption is that semantic relationships, such as hypernym, can be systematically expressed through lexicalsyntactic patterns. Starting with some of these patterns, originally described in Hearst (1992), I (i) identify new patterns that express hypernym; (ii) adapt three other patterns (Hearst, 1992), considering specificities of the Portuguese language; and (iii) intersect these results, in order to produce inferences. The theoretical approach is inspired by the wittgensteinian ideas about meaning. The resulting ontology´s most prominent features are: (i) the fact that it does not have a priori categories, since categories are human constructs, abstractions that reflect a particular world view. Instead of discussing what should be the main categories in a domain, sustaining the ontology on corpora assumes that the corpus reflects the implicit knowledge of a given domain; and (ii) the fact that it does not have a priori definitions: the meaning of a word is derived from its relations with other words. The method - automatic extraction of semantic relations through rules, and the intersection of this information in order to produce inferences - was applied to two corpora: a health domain corpus and a generic corpus. The positive results show that the method can be very useful in ontology building and it can also be a valuable tool for lexicographers and named entity recognition systems. The main advantages of the method are (i) the simplicity of automating the process of ontology building; (ii) the ease of categorizing specialized domains, and (iii) its dynamicity, since the possibility of constantly updating the corpus makes it less subject to errors.