this is a problem of great interest to us, and to others in the so-called 'ontology learning' area (a bit of a misnomer). Besides the pointers already suggested by Katrin Erk, a survey of recent work in the area can be found in the recent book edited by Buitelaar and Cimiano ("Ontology Learning and Population: Bridging the Gap between Text and Knowledge", published by IOS press). For the literature before around 2003, see also the web page of a graduate seminar we had in Essex a few years back:
In our own work, we have been primarily interested in how to extract attributes and values of these attributes from text. Abdulrahman Almuhareb's 2006 Essex Computer Science dissertation, "Attributes in Lexical Acquisition", proposes unsupervised and supervised methods for identifying concept attributes and values using relation extraction techniques, and also some preliminary work on relating values to attributes using patterns. In ongoing work at the University of Trento, we have been trying with Marco Baroni, Eduard Barbu, and other colleagues to improve the quality of the extracted attributes (e.g., Marco Baroni's 'Strudel' model) and to develop better evaluation methods relying both on the feature norms collected by psychologists and on techniques for extracting such values and attributes from WordNet (Eduard Barbu just presented a paper on this topic at the Global WordNet conference).