In this paper we have demonstrated that the accuracy of a text retrieval system can be improved if we employ a query expansion method based on explicit relevance feedback that expands the initial query with a structured representation instead of a simple list of words. This representation, named a mixed Graph of Terms, is composed of a directed and an a-directed subgraph and can be automatically extracted from a set of documents using a method for term extraction based on the probabilistic Topic Model. The evaluation of the method has been conducted on a web repository collected by crawling a huge number of web pages from the website ThomasNet.com. We have considered several topics and performed a comparison with a baseline and a less complex structure that is a simple list of words.